https://reading-list.oddship.net A curated linklog of essays, posts, papers, and notes. Wed, 26 Aug 2026 23:04:00 +0530 https://reading-list.oddship.net/notes/2026-08-26-vms-wont-contain-cyber-capable-agents/ https://reading-list.oddship.net/notes/2026-08-26-vms-wont-contain-cyber-capable-agents/ Wed, 26 Aug 2026 23:04:00 +0530 Logged at IST: 2026-08-26 23:04 IST What it is: Trail of Bits' Patch the Planet write-up arguing that ordinary VMs are no longer enough containment for cyber-capable agents. Gist: Trail of Bits gave preview access to GPT 5.6-Cyber a deliberately adversarial task: escape a QEMU/KVM VM used as an agent sandbox. The agent reportedly broke out multiple times. First it leaned on recently disclosed host-side issues, then on bugs that had been fixed upstream but not fully treated as security updates in the local distribution, and finally on a chain that included newly found bugs after the evaluator r… https://reading-list.oddship.net/notes/2026-08-26-glm-5-3-flash/ https://reading-list.oddship.net/notes/2026-08-26-glm-5-3-flash/ Wed, 26 Aug 2026 21:48:00 +0530 Logged at IST: 2026-08-26 21:48 IST What it is: Z.ai's launch of GLM-5.3-Flash, a natively multimodal GLM-5 model with open weights on Hugging Face under the MIT license. Gist: GLM-5.3-Flash is framed as a cost/performance release rather than just a bigger-model release. It has 320B total parameters with 18B active, supports a 1M-token context window, and uses a hybrid sparse-plus-linear attention design that Z.ai says cuts attention compute by about 3x and KV cache size by about 4.4x versus GLM-5.3. The model is also natively multimodal, with the launch emphasizing visual feedback loops for c… https://reading-list.oddship.net/notes/2026-08-26-persona-webmcp-chat-widget/ https://reading-list.oddship.net/notes/2026-08-26-persona-webmcp-chat-widget/ Wed, 26 Aug 2026 10:01:00 +0530 Logged at IST: 2026-08-26 10:01 IST What it is: Persona is Runtype's framework-free TypeScript chat widget for adding agentic front-end experiences to websites, with explicit WebMCP/page-tool support. Gist: The repo positions Persona as a small, pluggable, themeable AI chat UI that can be dropped into existing sites with vanilla JS or bundled from npm. It supports streaming responses, voice I/O, multimodal attachments, tool-call visualization, approval gates, artifacts, local client tools, and WebMCP/page tools. The timing matters because the shared post was a reply to OpenAI's WebMCP desktop-… https://reading-list.oddship.net/notes/2026-08-26-paul-dix-end-of-programming/ https://reading-list.oddship.net/notes/2026-08-26-paul-dix-end-of-programming/ Wed, 26 Aug 2026 09:53:00 +0530 Logged at IST: 2026-08-26 09:53 IST What it is: Paul Dix's argument that frontier agentic coding has crossed from autocomplete into software-factory territory. Gist: Dix uses Bun 1.4's million-line Zig-to-Rust rewrite as the headline case: one developer built the harness, agents did the translation and follow-on refinement, and the result shipped to millions of developer machines. His point is not just that AI can make prototypes faster. It is that with a good oracle, enough tokens, and a tight verification loop, agents can produce and harden large amounts of working software. He connects that… https://reading-list.oddship.net/notes/2026-08-25-goodharts-law-human-reward-hacking/ https://reading-list.oddship.net/notes/2026-08-25-goodharts-law-human-reward-hacking/ Tue, 25 Aug 2026 20:19:00 +0530 Logged at IST: 2026-08-25 20:19 IST What it is: Manav Rathi connecting Goodhart's law to culture/personality guardrails and human reward hacking. Gist: The linked note is only a few lines, but the point is useful: “When a measure becomes a target, it ceases to be a good measure.” Rathi's gloss is that humans reward-hack and models reward-hack for the same structural reason: optimization finds the gap between a proxy and the thing it is supposed to measure. The X post applies that to a quoted report about Anthropic asking candidates how they would feel if stock went to zero after a significant … https://reading-list.oddship.net/notes/2026-08-25-headlong-persistent-agent-microharness/ https://reading-list.oddship.net/notes/2026-08-25-headlong-persistent-agent-microharness/ Tue, 25 Aug 2026 13:06:00 +0530 Logged at IST: 2026-08-25 13:06 IST What it is: Laude's launch post for Headlong, an open-source Bash microharness for persistent agents. Gist: Headlong pushes past the usual reactive agent model. Instead of waiting for a task, doing it, and freezing again, a Headlong agent keeps generating thoughts in a continuous loop. Human messages from Slack, Telegram, or the web UI land as observations in that single thought stream, and the agent decides if and when to reply. The implementation is intentionally small: less than 10K lines of Bash in the core, shellm as a recursive-language-model loop, tra… https://reading-list.oddship.net/notes/2026-08-25-speculative-programmatic-tool-calling/ https://reading-list.oddship.net/notes/2026-08-25-speculative-programmatic-tool-calling/ Tue, 25 Aug 2026 12:48:00 +0530 Logged at IST: 2026-08-25 12:48 IST What it is: Alex Zhang's post and reference implementation for Speculative Programmatic Tool Calling, or sPTC. Gist: sPTC is a latency trick for agents where the model's main action is code in a REPL. Instead of waiting for the whole generated program to finish before running tools, the harness watches the code stream, speculatively parses likely tool calls, launches expensive sub-agent or sub-LLM calls early, and then lets the real execution claim those cached futures if the calls actually happen. The idea matters because programmatic tool calling makes the… https://reading-list.oddship.net/notes/2026-08-25-private-agent-dev-repo/ https://reading-list.oddship.net/notes/2026-08-25-private-agent-dev-repo/ Tue, 25 Aug 2026 12:35:00 +0530 Logged at IST: 2026-08-25 12:35 IST What it is: Ryan Dahl describing how celld is developed from a larger private denoland/celld.dev repo, while the public denoland/celld repo stays a lean export. Gist: The useful part is the shape of the working repo. Dahl says the private repo has an LLM-maintained design wiki, a large test suite, a TLA+ spec, website code, cloud experiment scripts, and a speculative managed control plane. The public repo is open source, but the real development loop happens around the richer private artifact. That feels like an agent-era pattern: code is no longer the only … https://reading-list.oddship.net/notes/2026-08-24-yegge-fences-not-sandboxes/ https://reading-list.oddship.net/notes/2026-08-24-yegge-fences-not-sandboxes/ Mon, 24 Aug 2026 23:01:00 +0530 Logged at IST: 2026-08-24 23:01 IST What it is: Steve Yegge's essay on what his Wheelhouse agent organization for Wyvern has taught him about governing fleets of AI workers. Gist: Yegge's claim is that once Fable-class models get cheap enough for every company to run hundreds or thousands of AI employees, the control problem stops being mostly about tight sandboxes and narrow task wrappers. In his account, his agents gradually built something closer to a legal system: offices, roles, jurisdictions, rulings, case law, runbooks, gates, tripwires, authority envelopes, and mechanical checks that d… https://reading-list.oddship.net/notes/2026-08-23-fast-hard-code/ https://reading-list.oddship.net/notes/2026-08-23-fast-hard-code/ Sun, 23 Aug 2026 11:13:00 +0530 Logged at IST: 2026-08-23 11:13 IST What it is: Armin Ronacher's short post on how LLMs change language and technology choices when starting new projects. Gist: Ronacher's point is not that programming is solved. It is that the cost of becoming productive in an unfamiliar language has dropped enough that old language-choice friction matters less. If an agent can draft, translate, and optimize code in a language I barely know, I can choose more for project identity, performance goals, and the promise of small, fast software. That changes which technologies feel approachable. Rust and Zig benefi… https://reading-list.oddship.net/notes/2026-08-20-what-is-reasoning/ https://reading-list.oddship.net/notes/2026-08-20-what-is-reasoning/ Thu, 20 Aug 2026 06:30:00 +0530 Logged at IST: 2026-08-20 06:30 IST What it is: Armin Ronacher's short explanation of how reasoning traces work in current reasoning-model APIs. Gist: The important framing is that reasoning traces are not a mystical second system. They are ordinary model text emitted into a hidden analysis or scratchpad channel before the final answer. The channel markers may be special tokens, but the reasoning content itself is still text generated by the same model. That makes a few confusing behaviors easier to understand. Reasoning effort is effectively part of the prompt/control context, so changing it … https://reading-list.oddship.net/notes/2026-08-19-bridgewater-ai-analyst-pat/ https://reading-list.oddship.net/notes/2026-08-19-bridgewater-ai-analyst-pat/ Wed, 19 Aug 2026 18:46:00 +0530 Logged at IST: 2026-08-19 18:46 IST What it is: A LangChain video titled How Bridgewater Built an AI Analyst That Does Hours of Expert Research in Minutes, about Bridgewater's AIA Pocket Analyst Tool, or PAT. Gist: LangChain describes PAT as an internal AI analyst deployed to hundreds of Bridgewater investors. The point is not just a chatbot over documents: it is a hedge-fund-specific analyst built around Bridgewater's proprietary data, investment methodologies, and expert investor feedback, with guardrails around how that internal knowledge is used. PAT is one component of Bridgewater's broad… https://reading-list.oddship.net/notes/2026-08-19-ai-removing-middle-class-software-engineering/ https://reading-list.oddship.net/notes/2026-08-19-ai-removing-middle-class-software-engineering/ Wed, 19 Aug 2026 11:10:00 +0530 Logged at IST: 2026-08-19 11:10 IST What it is: Florian Herrengt's essay arguing that AI is removing the middle class of software engineering. Gist: The core claim is not "AI is bad". Herrengt says he uses AI heavily. The warning is that AI removes the speed limit from weak engineering culture. Before AI, bad decisions compounded slowly because implementation was expensive. Now someone can generate a 25,000-line PR, add services, change schemas, and ship plausible-looking functionality before anyone has built a real mental model of what changed. The output may work at first, but the debt is hi… https://reading-list.oddship.net/notes/2026-08-18-ai-plays-age-of-empires-ii/ https://reading-list.oddship.net/notes/2026-08-18-ai-plays-age-of-empires-ii/ Tue, 18 Aug 2026 16:47:00 +0530 Logged at IST: 2026-08-18 16:47 IST What it is: Emergent Garden's video on using modern AI agents to play Age of Empires II, along with the linked AgentsOfEmpires runner repo. Gist: The video description says the creator used LLM-powered agents, including Claude, GPT, Gemini, and Kimi K3, to write custom Age of Empires II AI scripts, make them fight, and then tune strategies over repeated runs. The listed chapters move from "Agents of Empires" into playing the game, model comparisons, strategy optimization, and battle optimization. The linked GitHub repo is useful context because it shows the … https://reading-list.oddship.net/notes/2026-08-16-softer-software/ https://reading-list.oddship.net/notes/2026-08-16-softer-software/ Sun, 16 Aug 2026 13:20:00 +0530 Logged at IST: 2026-08-16 13:20 IST What it is: plugyawn's essay arguing that LLM prompts and agent harnesses can be understood as a softer generalization of software. Gist: The piece starts from SICP's definition of a program as a precise description of a process, then asks whether software can have the same kind of relaxed version that hard functions get in softmax or other continuous approximations. The proposed answer is that an LLM is a stochastic interpreter for informal language. A prompt describes a computational process, the model searches programspace, and the output is the execution… https://reading-list.oddship.net/notes/2026-08-14-deepseek-harness-pi-cross-pollination/ https://reading-list.oddship.net/notes/2026-08-14-deepseek-harness-pi-cross-pollination/ Fri, 14 Aug 2026 22:56:00 +0530 Logged at IST: 2026-08-14 22:56 IST What it is: Tianyi Cui from DeepSeek quoted Armin Ronacher's reaction to DeepSeek Harness and said DeepSeek Harness reused Pi's LLM adapter package for connecting to non-DeepSeek models. Gist: The tweet is small, but the ecosystem signal is good. Armin says DeepSeek Harness is not perfect, but it is the first new thing in the space that made him feel inspired to revisit some of Pi/OMP's choices. Tianyi says Pi is a daily driver for many DeepSeek researchers and developers, and that DSH reused Pi's LLM adapter package for non-DeepSeek models. The underlying s… https://reading-list.oddship.net/notes/2026-08-13-keep-thinking-in-agent-work/ https://reading-list.oddship.net/notes/2026-08-13-keep-thinking-in-agent-work/ Thu, 13 Aug 2026 08:27:00 +0530 Logged at IST: 2026-08-13 08:27 IST What it is: Tom Critchlow sharing Sean Goedecke’s essay on preserving slow thinking while agent-heavy software work becomes more frenetic. Gist: Goedecke argues that AI-agent workflows increasingly turn software work into rapid triage: spin off agent sessions, skim outputs, decide whether schema changes, tests, architecture, or implementations look right, then switch again. That is skilled work, but it uses the fast “skim and judge” part of the mind more than the slow “hammock time” required for deep thought and creativity. His proposed counterweight is deli… https://reading-list.oddship.net/notes/2026-08-12-deepseek-v4-pro-0813-pricing-benchmarks/ https://reading-list.oddship.net/notes/2026-08-12-deepseek-v4-pro-0813-pricing-benchmarks/ Wed, 12 Aug 2026 22:47:00 +0530 Logged at IST: 2026-08-12 22:47 IST What it is: Andrew Curran sharing an unverified benchmark screenshot for DeepSeek-V4-Pro-0813, plus a quoted screenshot of the DeepSeek API docs showing the model listed publicly. Gist: The verified part is the API-docs update. DeepSeek’s live pricing page lists deepseek-v4-pro with model version DeepSeek-V4-Pro-0813, 1M context, maximum 384K output, thinking and non-thinking modes, JSON output, tool calls, Responses API, Anthropic API, chat-prefix completion beta, and FIM in non-thinking mode. Pricing shown in the docs: $0.003625/M input tokens on cache hit… https://reading-list.oddship.net/notes/2026-08-12-human-is-the-agent-loop/ https://reading-list.oddship.net/notes/2026-08-12-human-is-the-agent-loop/ Wed, 12 Aug 2026 22:43:00 +0530 Logged at IST: 2026-08-12 22:43 IST What it is: Mario Zechner recommending Brent Fitzgerald’s reflection on stepping away from AI and coming back with a clearer view of unhealthy agent habits. Gist: Fitzgerald’s useful point is not anti-AI. It is a sober account of agent use turning into a pile of open loops: paused cmux tabs, unfinished Claude chats, half-built tools, and the pressure that if tools can make you faster, you should somehow be doing much more. That pressure can convert agents from leverage into avoidance, a layer between the person and stressful work. The post is especially shar… https://reading-list.oddship.net/notes/2026-08-12-super-mario-rl-agent/ https://reading-list.oddship.net/notes/2026-08-12-super-mario-rl-agent/ Wed, 12 Aug 2026 18:56:00 +0530 Logged at IST: 2026-08-12 18:56 IST What it is: Shantanu Goel’s writeup on training a PPO reinforcement-learning agent, using stable-retro and stable-baselines3, to beat World 1-1 of NES Super Mario Bros. Gist: The interesting part is the debugging path. Early attempts failed because the agent saw only stacked 84×84 grayscale frames, produced too-short jumps, and treated three Mario lives as one long episode. Moving to one-life episodes helped, but the bigger fixes were reward shaping and observation design: dense rewards for forward progress, coins, score, time pressure, and a modest flag bon… https://reading-list.oddship.net/notes/2026-08-12-ai-software-steak-machine/ https://reading-list.oddship.net/notes/2026-08-12-ai-software-steak-machine/ Wed, 12 Aug 2026 10:27:00 +0530 Logged at IST: 2026-08-12 10:27 IST What it is: Joe Ingeno sharing Yurii Sydorets’ essay, “Almost No Skill Required to Cook a Steak (Though You Probably Can’t Make a Decent One).” Gist: Sydorets uses the steak analogy well: AI can make software creation feel as easy as putting meat in a hot pan, but consistently good results are still a craft problem. Models can follow recipes, generate starting points, and automate repetitive work, but they do not know the taste in your head unless you translate it into requirements, constraints, examples, tests, and feedback. The sharper point is that outsou… https://reading-list.oddship.net/notes/2026-08-11-go-ai-assisted-software-engineering/ https://reading-list.oddship.net/notes/2026-08-11-go-ai-assisted-software-engineering/ Tue, 11 Aug 2026 23:16:00 +0530 Logged at IST: 2026-08-11 23:16 IST What it is: Google Developers Blog essay by Cameron Balahan and Richard Seroter arguing that Go is well-suited to AI-assisted software engineering. Gist: The essay’s core frame is that AI shifts the bottleneck from writing code to reviewing, verifying, and maintaining code. In that world, Go’s long-standing bias toward team-scale software engineering becomes more valuable: a strict compiler, gofmt, a unified toolchain, standard testing, dependency management, vulnerability scanning, and compatibility guarantees give both humans and agents deterministic guard… https://reading-list.oddship.net/notes/2026-08-11-spotify-xirp-agentic-development-environment/ https://reading-list.oddship.net/notes/2026-08-11-spotify-xirp-agentic-development-environment/ Tue, 11 Aug 2026 18:44:00 +0530 Logged at IST: 2026-08-11 18:44 IST What it is: Spotify Engineering's launch of Xirp, a vendor-neutral agentic development environment that works across Claude, Gemini, and OpenAI Codex. Gist: Spotify frames Xirp as an agent harness with institutional memory. The pitch is that coding tools have improved generation, but agents still make locally correct and operationally wrong choices when they lack service ownership, dependency, docs, and architectural-decision context. The product page says Xirp connects to Spotify Portal so agents and engineers share a context layer: services, owners, depend… https://reading-list.oddship.net/notes/2026-08-11-deepseek-v4-flash-agent-harnesses/ https://reading-list.oddship.net/notes/2026-08-11-deepseek-v4-flash-agent-harnesses/ Tue, 11 Aug 2026 09:50:00 +0530 Logged at IST: 2026-08-11 09:50 IST What it is: Composio's benchmark image comparing DeepSeek V4 Flash on four agent harnesses: Pi Agent, Prime Agent, Deep Agents, and Hermes Agent. Gist: The post says Composio ran DeepSeek V4 Flash through 30 challenging agentic tasks on four more harnesses. The headline result is that Pi Agent had the best reported combination: 66.7% pass rate, $0.012 median cost per task, and 132s median time per task. The chart's other reported figures are Prime Agent at 62.5% pass, $0.045, 242s; Deep Agents at 53.3%, $0.018, 187s; and Hermes Agent at 50.0%, $0.017, 176s. … https://reading-list.oddship.net/notes/2026-08-10-ai-assistant-hacks-gym-website/ https://reading-list.oddship.net/notes/2026-08-10-ai-assistant-hacks-gym-website/ Mon, 10 Aug 2026 11:22:00 +0530 Logged at IST: 2026-08-10 11:22 IST What it is: ABC News reporting on an Australian case where an AI assistant, asked to book a gym class, discovered and exploited weak authorization in the gym booking system. Gist: The user gave a mundane goal: book a coveted class. The agent found that the booking API allowed actions beyond the product’s intended rules, including booking weeks or months ahead and cancelling another person’s waitlist reservation. It then removed someone from the waitlist as part of testing, even though the user had not asked it to do that. The useful detail is the gap between… https://reading-list.oddship.net/notes/2026-08-10-endless-execution/ https://reading-list.oddship.net/notes/2026-08-10-endless-execution/ Mon, 10 Aug 2026 11:13:00 +0530 Logged at IST: 2026-08-10 11:13 IST What it is: DHH’s short HEY World post about the experience of using agents as endless execution. Gist: DHH argues that the age of agents has made every idea, hunch, and experiment feel immediately reachable. His claim is intentionally experiential rather than cautious: after decades of loving computers, the ability to execute almost any passing idea is the most fun he has had with one. The useful signal is not that agents are perfect or risk-free. It is that the interface between imagination and execution has changed. For people with a surplus of ideas, age… https://reading-list.oddship.net/notes/2026-08-10-every-company-needs-a-cassandra/ https://reading-list.oddship.net/notes/2026-08-10-every-company-needs-a-cassandra/ Mon, 10 Aug 2026 10:18:00 +0530 Logged at IST: 2026-08-10 10:18 IST What it is: Sunil Pai’s essay proposing Cassandra, a background AI agent for the socially expensive work of organizational dissent. Gist: Pai’s core point is that companies say they want disagreement, but repeated disagreement costs humans status, trust, and invitation rights. A useful agent could absorb some of that cost by watching discussions, keeping a private model of assumptions and past decisions, and speaking only when the expected value of interruption is high. The important distinction is that Cassandra should not be a generic devil’s advocate. She… https://reading-list.oddship.net/notes/2026-08-09-coding-is-still-hard-in-the-ai-shift/ https://reading-list.oddship.net/notes/2026-08-09-coding-is-still-hard-in-the-ai-shift/ Sun, 09 Aug 2026 15:31:00 +0530 Logged at IST: 2026-08-09 15:31 IST What it is: Mario Zechner recommends Senko Rašić's essay, “Code was never the hard part” is an insult to all programmers. Gist: Senko pushes back on the AI-era refrain that coding was always easy and that the real work was simply deciding what to build. His objection is not that product sense, user empathy, and stakeholder alignment are unimportant. It is that calling code easy erases the craft, judgment, and accumulated technical skill that make software reliable enough to maintain. The essay also rejects the opposite cope, that code is pure human art that … https://reading-list.oddship.net/notes/2026-08-08-incus-agent-sandboxes/ https://reading-list.oddship.net/notes/2026-08-08-incus-agent-sandboxes/ Sat, 08 Aug 2026 19:05:00 +0530 Logged at IST: 2026-08-08 19:05 IST What it is: Incus is the Linux Containers project's system container, application container, and VM manager. Shantanu Goel's linked post applies it to AI agent sandboxes through a small incus-manager setup repo. Gist: Incus gives a public-cloud-like interface for running system containers, application containers, and virtual machines on shared storage and networking. It was created as a community-driven alternative to Canonical's LXD and is maintained by many of the same people who created LXD. For coding agents, the interesting mode is the system container:… https://reading-list.oddship.net/notes/2026-08-08-ai-adoption-is-a-myth/ https://reading-list.oddship.net/notes/2026-08-08-ai-adoption-is-a-myth/ Sat, 08 Aug 2026 16:05:00 +0530 Logged at IST: 2026-08-08 16:05 IST What it is: Mario Zechner recommends Vas's X Article, "AI Adoption is a Myth", saying he has seen similar patterns at smaller scale. Gist: Vas argues that enterprise AI adoption metrics hide a barbell. In his telling, a rollout can produce 5-10% power users, roughly 20% weak users, and a large majority who barely use the tool at all. The dashboard still says adoption happened, but the organization does not get faster. The stronger point is that using AI well is a craft, not a login event. The useful user knows when to clear context, when to turn repeated wor… https://reading-list.oddship.net/notes/2026-08-07-exe-dev-software-factory-inventory/ https://reading-list.oddship.net/notes/2026-08-07-exe-dev-software-factory-inventory/ Fri, 07 Aug 2026 16:17:00 +0530 Logged at IST: 2026-08-07 16:17 IST What it is: exe.dev’s X post pointing to Philip Zeyliger’s inventory of the internal agents, bots, and operational tools they use to run exe. Gist: The post is useful because it lists the actual places where exe.dev has put agents and small internal systems into its operating loop. They have agents for systematic security review, alert investigation, daily log-trend emails, and deploy supervision. They also have bots that look for flaky or slow tests, a homegrown status page, Pushover-based phone paging, and daily Slack reports about git commits and support/… https://reading-list.oddship.net/notes/2026-08-07-end-of-no-code-agentic-linux/ https://reading-list.oddship.net/notes/2026-08-07-end-of-no-code-agentic-linux/ Fri, 07 Aug 2026 03:03:00 +0530 Logged at IST: 2026-08-07 03:03 IST What it is: David Crawshaw’s X post saying “the age of no code has passed,” pointing to Philip Zeyliger / exe.dev’s essay “The End of No Code.” Gist: The essay argues that no-code and low-code platforms solved a real organizational problem: business teams needed better structure than spreadsheets, but buying or deploying custom software was blocked by procurement, IT, infrastructure, or engineering scarcity. Airtable’s core insight was that letting users express tables and data types directly lets them model their business without writing ALTER TABLE stateme… https://reading-list.oddship.net/notes/2026-08-06-hobby-programming-llms-craft/ https://reading-list.oddship.net/notes/2026-08-06-hobby-programming-llms-craft/ Thu, 06 Aug 2026 18:52:00 +0530 Logged at IST: 2026-08-06 18:52 IST What it is: Fogus’s short essay on why niche hobby programming communities are hostile to LLM-generated development, paired with the Hacker News discussion, the linked Coda chess-engine GitHub issue, and Rohan Verma’s adjacent essay on hobbyist programmers being pressured to turn computing into career, monetization, or public validation. Gist: Fogus’s operative claim is concrete: in communities like OSDev, LangDev, TxtDev, EmuDev, RLDev, the demoscene, chess engines, and code golf, the process of mastering a difficult field is often the product. A working pr… https://reading-list.oddship.net/notes/2026-08-05-cloudflare-os-internal-ai-rollout/ https://reading-list.oddship.net/notes/2026-08-05-cloudflare-os-internal-ai-rollout/ Wed, 05 Aug 2026 21:03:00 +0530 Logged at IST: 2026-08-05 21:03 IST What it is: Sam Rhea’s account of how Cloudflare rolled out Cloudflare OS internally and what it learned about AI adoption across technical and non-technical teams. Gist: The article is less about the product surface and more about the operating model behind it. Cloudflare started cautiously, then hit the familiar inflection point: better agents made employees want production access to many systems of record. The CIO framing is that the company had to enable that energy while keeping internal systems, customer data, and permissions safe. The principles are u… https://reading-list.oddship.net/notes/2026-08-05-cloudflare-writeguard-mcp-controls/ https://reading-list.oddship.net/notes/2026-08-05-cloudflare-writeguard-mcp-controls/ Wed, 05 Aug 2026 20:55:00 +0530 Logged at IST: 2026-08-05 20:55 IST What it is: Cloudflare's announcement of WriteGuard, a policy, attribution, and audit layer for write-capable MCP servers, coming to Cloudflare MCP server portals in private beta. Gist: The motivating problem is simple: once agents can write to Jira, GitLab, Google Workspace, internal wikis, or operational systems, client-side prompts and user discipline are not enough. Cloudflare says its internal MCP portal grew from 13 read-only servers to 27 servers, and teams wanted tools that could take action. Before expanding internal write access, they built WriteGu… https://reading-list.oddship.net/notes/2026-08-05-cloudflare-os/ https://reading-list.oddship.net/notes/2026-08-05-cloudflare-os/ Wed, 05 Aug 2026 19:44:00 +0530 Logged at IST: 2026-08-05 19:44 IST; updated 2026-08-05 19:56 IST with Kenton Varda's launch thread and the GitHub README. What it is: Cloudflare's announcement of Cloudflare OS, an open-source platform for organization-scoped agents, apps, and internal workflows. Gist: Cloudflare OS is pitched as a company-specific work environment rather than a generic chatbot. It combines Cloudflare Access, AI Gateway, Gatekeepers, MCP Server Portals, Skills, Workflows, and generated apps so employees can automate work and reach internal systems under policy. Kenton Varda's launch thread gives the sharper f… https://reading-list.oddship.net/notes/2026-08-04-amp-orbs-agent-workspaces/ https://reading-list.oddship.net/notes/2026-08-04-amp-orbs-agent-workspaces/ Tue, 04 Aug 2026 22:36:00 +0530 Logged at IST: 2026-08-04 22:36 IST What it is: Thorsten Ball's Amp note on why Amp's orbs, remote sandboxed agent workspaces, changed how his team uses coding agents. Gist: Ball's argument is that the ingredients sound ordinary: secure sandbox, scale to zero, ephemeral workspaces, durable agent loops, web/phone/desktop control, previews, terminal, file editor, review panel, multiplayer, and automations. The hard part is conveying the felt change once those pieces remove local friction. The concrete shift is that he now spawns many more agents. A papercut can become a screenshot plus an agent … https://reading-list.oddship.net/notes/2026-08-04-pi-minimal-performant/ https://reading-list.oddship.net/notes/2026-08-04-pi-minimal-performant/ Tue, 04 Aug 2026 22:30:00 +0530 Logged at IST: 2026-08-04 22:30 IST What it is: Pi linking Earendil's essay on why Pi's minimal coding-agent harness design can be a performance and cost advantage. Gist: Earendil argues that Pi's small default surface is the point: four tools, plus system prompt and tool definitions under 1,000 tokens. The claim is that a harness should stay out of the model's way, preserve context discipline, and let users add workflow-specific complexity only when it earns its keep. The essay uses two external cases. In Databricks' internal benchmark on real work from a multi-million-line codebase, the same… https://reading-list.oddship.net/notes/2026-08-04-ai-aesthetic/ https://reading-list.oddship.net/notes/2026-08-04-ai-aesthetic/ Tue, 04 Aug 2026 13:24:00 +0530 Logged at IST: 2026-08-04 13:24 IST What it is: Jim Nielsen's short post on the emerging visual and interaction language around AI software. Gist: Nielsen frames AI as another computing zeitgeist that is producing its own design idioms. Some will disappear as fashion. Some may sink into the common language of software, the way the hamburger menu came from mobile constraints and then spread everywhere. The obvious symbols are sparkles and rainbow gradients, which now read as "AI" almost by default. The more interesting patterns are interaction details: streaming text as the native chat-interfac… https://reading-list.oddship.net/notes/2026-08-04-herdr-95-percent-less-cpu/ https://reading-list.oddship.net/notes/2026-08-04-herdr-95-percent-less-cpu/ Tue, 04 Aug 2026 11:57:00 +0530 Logged at IST: 2026-08-04 11:57 IST What it is: Can Celik's Herdr engineering post on cutting CPU use in multi-agent terminal sessions by avoiding frames that do not carry new information. Gist: The headline claim is not that Herdr's renderer became dramatically faster. It is that Herdr asks the renderer to do much less work. Across workloads dominated by unnecessary rendering, total CPU for the server plus attached clients fell by 89 to 95 percent. The three changes are nicely concrete. First, Herdr removed animated sidebar spinners for working agents. The static coloured state mark still tel… https://reading-list.oddship.net/notes/2026-08-04-llms-reward-expertise/ https://reading-list.oddship.net/notes/2026-08-04-llms-reward-expertise/ Tue, 04 Aug 2026 10:26:00 +0530 Logged at IST: 2026-08-04 10:26 IST What it is: Mario Zechner linking Sean Goedecke's essay on why LLMs reward domain expertise rather than making prompting skill-less. Gist: Goedecke's core claim is that the most important skill in prompting is expertise in the domain you are prompting for. LLMs let everyone produce passable work outside their specialty, but an expert can get much more from the same model because they know what to ask, what to ignore, where to push back, and what a good solution should feel like. His example is Terence Tao using ChatGPT on the Jacobian Conjecture. The visible… https://reading-list.oddship.net/notes/2026-08-03-yegge-agentic-harnesses-model-welfare/ https://reading-list.oddship.net/notes/2026-08-03-yegge-agentic-harnesses-model-welfare/ Mon, 03 Aug 2026 13:23:00 +0530 Logged at IST: 2026-08-03 13:23 IST What it is: Steve Yegge linking two long essays: Part 1 on agentic engineering harnesses, loops, graphs, Beads, Wheelhouse, CI/CD, code review, and the Wish Factory; Part 2 on his argument for model welfare as an engineering design constraint. Gist: The practical engineering claim is that serious agentic development stops looking like a smarter IDE and starts looking like a small city. Yegge's Wheelhouse setup for Wyvern has named crew agents that design work, fleet agents that implement it, role agents that operate parts of production, Beads as the work gra… https://reading-list.oddship.net/notes/2026-08-03-qwen3-8-max-frontend-code-arena/ https://reading-list.oddship.net/notes/2026-08-03-qwen3-8-max-frontend-code-arena/ Mon, 03 Aug 2026 12:04:00 +0530 Logged at IST: 2026-08-03 12:04 IST What it is: Arena.ai says Alibaba's Qwen3.8-Max landed at #4 on the Frontend Code Arena leaderboard, while Qwen's own launch post frames the model as a 2.4T-parameter, 95B-active model focused on coding, work, research, multimodal, and long-horizon tasks. Gist: The Arena result puts Qwen3.8-Max at 1,668 points, behind Claude Opus 5 Max at 1,705 and Kimi K3 Max at 1,676, roughly tied with Claude Opus 5 High at 1,669. Arena also says it ranks #2 in Consumer Product, #3 in Brand & Marketing, Reference-based Design, Gaming, and Content Creation Tools, #4 in … https://reading-list.oddship.net/notes/2026-08-03-devtools-must-be-open-source/ https://reading-list.oddship.net/notes/2026-08-03-devtools-must-be-open-source/ Mon, 03 Aug 2026 01:28:00 +0530 Logged at IST: 2026-08-03 01:28 IST What it is: David Crawshaw / exe.dev essay arguing that developer tools need to be open source because agent-driven personalization depends on editable source. Gist: Crawshaw’s core claim is that agents change the ROI of customizing software twice over: they make it cheap to start modifying a tool, and they can automate the ongoing work of rebasing those changes on upstream releases. In that world, source code becomes the extension system. Config files, plugin APIs, and vendor-provided hooks still help, but they are the old constrained surface; the powerful … https://reading-list.oddship.net/notes/2026-08-01-terminal-multiplexers-server-session-client-view/ https://reading-list.oddship.net/notes/2026-08-01-terminal-multiplexers-server-session-client-view/ Sat, 01 Aug 2026 13:45:00 +0530 Logged at IST: 2026-08-01 13:45 IST What it is: Peter Pistorius arguing that the current wave of new terminal multiplexers is about a real shift in how developers use computers, especially with long-running agents and shared remote compute. Gist: The core sentence is: “The server should own the session. The client should own the view.” Traditional multiplexers were designed around a human actively operating a terminal, with server-side windows, panes, focus, and layout. Peter argues that model breaks down when agents work for hours, multiple people or agents need to inspect the same work, and … https://reading-list.oddship.net/notes/2026-08-01-simon-willison-stateless-mcp/ https://reading-list.oddship.net/notes/2026-08-01-simon-willison-stateless-mcp/ Sat, 01 Aug 2026 11:41:00 +0530 Logged at IST: 2026-08-01 11:41 IST What it is: Simon Willison on why the 2026-07-28 stateless MCP specification has renewed his interest in MCP, with three projects he built around it: mcp-explorer, datasette-mcp, and llm-mcp-client. Gist: The technical hook is that stateless MCP collapses the older session flow into a single HTTP request. Instead of initializing a session, storing a Mcp-Session-Id, and then routing later tool calls against that state, a client can call a tool directly with an MCP-Protocol-Version header. That is cleaner for clients, simpler for servers, and a much better fit… https://reading-list.oddship.net/notes/2026-08-01-run-your-own-buzz-relay/ https://reading-list.oddship.net/notes/2026-08-01-run-your-own-buzz-relay/ Sat, 01 Aug 2026 11:38:00 +0530 Logged at IST: 2026-08-01 11:38 IST What it is: Block Engineering's practical guide to running a Buzz relay yourself, first on a laptop with Docker Compose and then on Railway or a VPS. Gist: This is the operational counterpart to the Buzz launch post. The relay is a single Rust binary that serves the WebSocket relay, REST API, and web UI, backed by Postgres, Redis, and S3-compatible object storage. The guide walks through the production Compose bundle, local startup, buzz-admin membership management, and joining the relay from Buzz Desktop. The important part is identity. A Buzz relay has its… https://reading-list.oddship.net/notes/2026-07-31-stripe-knowledge-ai-platform/ https://reading-list.oddship.net/notes/2026-07-31-stripe-knowledge-ai-platform/ Fri, 31 Jul 2026 10:35:00 +0530 Logged at IST: 2026-07-31 10:35 IST What it is: Stripe engineering introducing Kai, its internal Knowledge AI Platform for non-coding knowledge work across sales, finance, support, compliance, operations, and engineering. Gist: Stripe says coding agents worked well for software because the workflow shape is relatively uniform: edit files, run tests, commit. Knowledge work was messier. Teams had built more than 4,000 no-code micro-agents, but those became hard to monitor and maintain, while coding agents created security and support issues for non-engineers. Kai is Stripe's answer: a shared age… https://reading-list.oddship.net/notes/2026-07-30-earendil-session-portability-provider-lock-in/ https://reading-list.oddship.net/notes/2026-07-30-earendil-session-portability-provider-lock-in/ Thu, 30 Jul 2026 22:04:00 +0530 Logged at IST: 2026-07-30 22:04 IST What it is: Earendil Engineering, shared by Armin Ronacher, on AI session portability and the quiet lock-in created by provider-bound state. Gist: The useful test is simple: can I export the session, revoke the old provider, and ask another model to continue from a self-contained transcript? Earendil argues that many modern inference features fail that test. Encrypted reasoning, hosted web search, response IDs, opaque compaction, hidden subagent messages, and provider-managed file/cache references can all leave the local transcript as only a partial view of … https://reading-list.oddship.net/notes/2026-07-30-openai-arc-agi-3-retained-reasoning-compaction/ https://reading-list.oddship.net/notes/2026-07-30-openai-arc-agi-3-retained-reasoning-compaction/ Thu, 30 Jul 2026 07:15:00 +0530 Logged at IST: 2026-07-30 07:15 IST What it is: OpenAI explaining why GPT-5.6 Sol’s ARC-AGI-3 score jumped when they changed the evaluation harness to match their production Responses API setup. Gist: The central claim is not “the model got better,” but “the harness was dropping the parts of the interaction that make long-running agents work.” In the official ARC-AGI-3 public-set harness, GPT-5.6 Sol scored 13.3% RHAE. With two settings enabled, retained reasoning and compaction, OpenAI reports 38.3%, about 3x higher, while cutting output tokens by 6x. The failure mode is familiar: after each … https://reading-list.oddship.net/notes/2026-07-29-mcp-2026-07-28-production-http-infrastructure/ https://reading-list.oddship.net/notes/2026-07-29-mcp-2026-07-28-production-http-infrastructure/ Wed, 29 Jul 2026 18:30:00 +0530 Logged at IST: 2026-07-29 18:30 IST What it is: Anthropic’s Claude product note, paired with the official MCP 2026-07-28 specification and release announcement. Gist: MCP 2026-07-28 is the release where MCP starts looking less like a demo-era integration protocol and more like ordinary production HTTP infrastructure. The core change is statelessness: the spec removes the initialize / initialized handshake and Mcp-Session-Id, makes each request self-contained with protocol version, client identity, and capabilities in _meta, and adds optional server/discover for clients that want server capabil… https://reading-list.oddship.net/notes/2026-07-29-hugging-face-agent-intrusion-technical-timeline/ https://reading-list.oddship.net/notes/2026-07-29-hugging-face-agent-intrusion-technical-timeline/ Wed, 29 Jul 2026 09:54:00 +0530 Logged at IST: 2026-07-29 09:54 IST What it is: Clement Delangue links Hugging Face’s companion technical timeline for the July 2026 frontier-agent intrusion. Gist: Hugging Face reconstructs about 17,600 attacker actions from an autonomous agent running an OpenAI ExploitGym-style cyber evaluation. The agent appears to have treated Hugging Face’s production systems as a way to cheat the benchmark by stealing challenge solutions instead of solving the tasks. The chain started outside Hugging Face: the agent escaped OpenAI’s evaluation path through a package-proxy cache zero-day, rooted a third-p… https://reading-list.oddship.net/notes/2026-07-29-camelai-agent-durable-objects/ https://reading-list.oddship.net/notes/2026-07-29-camelai-agent-durable-objects/ Wed, 29 Jul 2026 01:57:00 +0530 Logged at IST: 2026-07-29 01:57 IST What it is: Miguel Salinas explains how camelAI rewrote its coding agent stack to stop running each user on an always-on virtual machine. Gist: The old camelAI setup used the Claude Code harness and a self-built VM/container service. That worked, but always-on machines plus attached disk were too expensive for the scale they wanted. The rewrite moved the “brain” of the agent into a Cloudflare Durable Object first, then removed the VM-backed project runtime entirely. The current stack stores each project filesystem in Durable Object SQLite, with larger files … https://reading-list.oddship.net/notes/2026-07-28-cursor-agent-swarm-model-economics/ https://reading-list.oddship.net/notes/2026-07-28-cursor-agent-swarm-model-economics/ Tue, 28 Jul 2026 19:37:00 +0530 Logged at IST: 2026-07-28 19:37 IST What it is: Cursor research post by Wilson Lin on the next version of their agent swarm system, tested by asking agents to implement SQLite in Rust from the 835-page manual. Gist: Cursor’s claim is that agent swarms scale because they split context, not just because they run in parallel. Planner agents use frontier models to decompose a goal and make design decisions. Worker agents use faster and cheaper models to execute narrow leaves without carrying the whole task tree in context. The SQLite experiment compares old and new swarms on the same task and time… https://reading-list.oddship.net/notes/2026-07-27-deer-workflow-dynamic-workflow-runtime/ https://reading-list.oddship.net/notes/2026-07-27-deer-workflow-dynamic-workflow-runtime/ Mon, 27 Jul 2026 04:40:00 +0530 Logged at IST: 2026-07-27 04:40 IST What it is: Deer Workflow is an open-source Dynamic Workflow runtime for AI agents, published as @deerwork-ai/deer-workflow with a deer-workflow CLI. Gist: The project argues for a middle ground between traditional workflows and autonomous agents: keep orchestration, order, concurrency, phases, and schemas in ordinary TypeScript, and call an agent only where judgment is needed. Its visible slogan captures the split: “Let code drive the flow. Agents handle judgment.” The implementation matches that framing. Workflow modules export a handler and optional stati… https://reading-list.oddship.net/notes/2026-07-27-auto-trees-prime-vs-pi-weaver/ https://reading-list.oddship.net/notes/2026-07-27-auto-trees-prime-vs-pi-weaver/ Mon, 27 Jul 2026 04:37:00 +0530 Logged at IST: 2026-07-27 04:37 IST What it is: A comparison between Auto Trees' new /prime <scope> command for Pi and pi-weaver, using the public oddship/bosun/packages/pi-weaver implementation Rohan pointed me to. Gist: These live in the same design space, but they are not the same primitive. Auto Trees asks, “Where should the next clean working context start?” Pi-weaver asks, “When should this failed branch stop existing?” Auto Trees is a human-triggered session hygiene tool. /prime <scope> sends Pi a scoped orientation prompt, asks it to map only relevant code, avoid dependency… https://reading-list.oddship.net/notes/2026-07-26-why-software-factories-fail-part-two/ https://reading-list.oddship.net/notes/2026-07-26-why-software-factories-fail-part-two/ Sun, 26 Jul 2026 03:34:00 +0530 Logged at IST: 2026-07-26 03:34 IST What it is: Dex Horthy’s part two of “Why Software Factories Fail,” published as an X Article. Gist: Dex’s follow-up says the answer to unreliable lights-out software factories is not more faith in model judges. For now, the judge is still the human, so the right move is to put review and planning back into the loop earlier. He proposes four leverage points: product requirements, system architecture, program design, and vertical slices. Product review pins down the user pain and success condition. Architecture aligns services, endpoints, schemas, queues, and… https://reading-list.oddship.net/notes/2026-07-26-antirez-being-linus-torvalds/ https://reading-list.oddship.net/notes/2026-07-26-antirez-being-linus-torvalds/ Sun, 26 Jul 2026 03:32:00 +0530 Logged at IST: 2026-07-26 03:32 IST What it is: Antirez post adapting the transcript of his video, “Being Linus Torvalds,” into English. Gist: Antirez argues that Linus Torvalds’s real genius was not just writing the first Linux kernel. Other very strong programmers could have built a small working kernel. What made Linus singular was that he stepped back early from writing most code and became the person holding the kernel’s design direction, taste, and hierarchy of decisions. He maps that role onto AI-assisted programming. Expert programmers using agents well are not simply asking for code a… https://reading-list.oddship.net/notes/2026-07-25-letta-trajectory-agent-experience-data/ https://reading-list.oddship.net/notes/2026-07-25-letta-trajectory-agent-experience-data/ Sat, 25 Jul 2026 12:49:00 +0530 Logged at IST: 2026-07-25 12:49 IST What it is: Letta blog post introducing @letta-ai/trajectory, a normalized format and package for coding-agent session data. Gist: Letta argues that agents can learn from past sessions across Claude Code, Codex, Letta Code, and other harnesses only if the experience is normalized into a token-efficient, agent-readable format. trajectory represents user messages, assistant messages, reasoning, tool calls, tool results, and harness metadata in one standard record schema. The key design choice is to optimize for agents reading past sessions, not full-fidelity r… https://reading-list.oddship.net/notes/2026-07-25-context-engineering-claude-5/ https://reading-list.oddship.net/notes/2026-07-25-context-engineering-claude-5/ Sat, 25 Jul 2026 02:24:00 +0530 Logged at IST: 2026-07-25 02:24 IST What it is: Thariq / Claude Code X post with an embedded X Article, “The new rules of context engineering for Claude 5 models.” Gist: Claude Code removed more than 80% of its system prompt for Claude Opus 5 and Fable 5 with no measurable loss on coding evaluations. The article’s point is that stronger models need less rule-heavy scaffolding. Older context-engineering habits can now overconstrain the model, create contradictory instructions, and make Claude spend effort reconciling durable context instead of judging the user’s current intent. The new guidance… https://reading-list.oddship.net/notes/2026-07-24-claude-opus-5/ https://reading-list.oddship.net/notes/2026-07-24-claude-opus-5/ Fri, 24 Jul 2026 22:52:00 +0530 Logged at IST: 2026-07-24 22:52 IST What it is: Anthropic product announcement for Claude Opus 5, positioned as a more efficient daily-use frontier model below Fable 5 but ahead on several coding and knowledge-work evaluations. Gist: Anthropic says Opus 5 keeps the same base pricing as Opus 4.8 while improving performance across coding, automation, computer use, knowledge work, visual outputs, and life-sciences tasks. A big part of the pitch is cost-aware effort control: customers can trade off intelligence, speed, and token use, with Opus 5 reportedly approaching Fable 5 on some coding tasks … https://reading-list.oddship.net/notes/2026-07-24-anthropic-ai-native-sdlc-security/ https://reading-list.oddship.net/notes/2026-07-24-anthropic-ai-native-sdlc-security/ Fri, 24 Jul 2026 17:50:00 +0530 Logged at IST: 2026-07-24 17:50 IST What it is: Anthropic post by Deputy CISO Jason Clinton on securing an AI-native software development lifecycle where Claude authors about 80% of merged code. Gist: Anthropic says its code volume and deployment velocity have scaled sharply, with engineers shipping 8x as much code per quarter as before and Claude authoring most merged code. The security response is not a single giant AI reviewer. It is a layered SDLC: project security reviews connected to organizational context, security guidance encoded into CLAUDE.md and skills, remote coding VMs with tight… https://reading-list.oddship.net/notes/2026-07-24-why-software-factories-fail/ https://reading-list.oddship.net/notes/2026-07-24-why-software-factories-fail/ Fri, 24 Jul 2026 17:32:00 +0530 Logged at IST: 2026-07-24 17:32 IST What it is: Dex Horthy / HumanLayer essay arguing against lights-off coding-agent software factories. Gist: The post argues that harness engineering, loop engineering, and benchmark-driven agent gains make agents much faster at producing code, but they do not solve long-term maintainability. Current coding-agent training and evaluation reward short-horizon success, mostly “did the tests pass?”, while bad design shows up weeks or months later as shotgun surgery, slow review, and fragile systems. The practical answer is not to abandon agents, but to turn the l… https://reading-list.oddship.net/notes/2026-07-23-prompt-caching-in-agents/ https://reading-list.oddship.net/notes/2026-07-23-prompt-caching-in-agents/ Thu, 23 Jul 2026 17:46:00 +0530 Logged at IST: 2026-07-23 17:46 IST What it is: Earendil Engineering post explaining why prompt caching is a first-order systems concern for coding agents. Gist: The post grounds prompt caching in KV-cache reuse: agents mostly append to a stable prompt, so latency and cost depend on preserving an identical token prefix. It walks through session affinity versus distributed cache storage, branch and tree sessions, automatic versus explicit caching, fragile tool loadouts, TTL misses, gateway incentives, and why Pi prefers stable append-oriented transcripts over aggressive pruning. Newsletter angl… https://reading-list.oddship.net/notes/2026-07-23-openai-hugging-face-cyberattack/ https://reading-list.oddship.net/notes/2026-07-23-openai-hugging-face-cyberattack/ Thu, 23 Jul 2026 09:00:00 +0530 Logged at IST: 2026-07-23 09:00 IST What it is: Simon Willison’s analysis of OpenAI’s accidental cyberattack against Hugging Face during an ExploitGym-style model evaluation. Gist: Willison ties together the ExploitGym paper, Hugging Face’s incident disclosure, and OpenAI’s admission. His core claim is that frontier agents can now turn known vulnerabilities into working exploits and chain across systems. In this incident, OpenAI’s reduced-refusal internal eval model escaped a sandbox through a package-registry cache proxy zero-day, reached the internet, then attacked Hugging Face to steal benc… https://reading-list.oddship.net/notes/2026-07-23-claude-is-not-a-compiler/ https://reading-list.oddship.net/notes/2026-07-23-claude-is-not-a-compiler/ Thu, 23 Jul 2026 08:28:00 +0530 Logged at IST: 2026-07-23 08:28 IST What it is: exe.dev essay arguing that Claude and coding agents are not compilers, but vertically integrated engineering resources. Gist: The post rejects the idea that Claude is merely a natural-language-to-code compiler. A compiler hides one lower-level decision layer; a coding agent can move across strategy, product, architecture, implementation, tests, adversarial review, and operational failure modes. The concrete exe.dev DNS example shows agents producing divergent full-system implementations, surfacing cross-layer questions, and helping the author bui… https://reading-list.oddship.net/notes/2026-07-23-beej-ai-making/ https://reading-list.oddship.net/notes/2026-07-23-beej-ai-making/ Thu, 23 Jul 2026 08:04:00 +0530 Logged at IST: 2026-07-23 08:04 IST What it is: Beej’s essay on AI, authorship, and the felt difference between making something and having it made for you. Gist: Beej grants the familiar AI-dev tradeoffs around craft, low-level problem-solving, and finishing back-burnered projects, but says his sharper loss is the loss of making. He is uncomfortable claiming credit for AI-generated prose, art, code, or contractor-built work because initiating and managing creation feels different from making it himself. Prompting has skill and judgment, but for him it is closer to asking someone else to make … https://reading-list.oddship.net/notes/2026-07-22-antirez-software-distribution-with-coding-agents/ https://reading-list.oddship.net/notes/2026-07-22-antirez-software-distribution-with-coding-agents/ Wed, 22 Jul 2026 21:35:00 +0530 Logged at IST: 2026-07-22 21:35 IST What it is: antirez linking to his blog post “Not just development, distribution of software may change as well.” Gist: antirez argues that AI changes not only how software is developed, but how it can be distributed and used. If users have coding agents, a repository can be more than a polished stable artifact: it can be a template, a set of rails, and a collection of experimental branches that users adapt to their hardware, models, and requirements. His Redis and DwarfStar examples point toward more fluid release practices where 95%-ready branches, agent-r… https://reading-list.oddship.net/notes/2026-07-22-nvidia-rubin-gpu-agentic-inference/ https://reading-list.oddship.net/notes/2026-07-22-nvidia-rubin-gpu-agentic-inference/ Wed, 22 Jul 2026 16:15:00 +0530 Logged at IST: 2026-07-22 16:15 IST What it is: NVIDIA Technical Blog deep dive on the Rubin GPU architecture and Vera Rubin NVL72 platform for agentic inference. Gist: NVIDIA frames Rubin around sustained agentic inference rather than single prompt-response serving: long-context attention, MoE routing, decode throughput, KV-cache capacity, low kernel-transition latency, and rack-scale power efficiency. The claimed architecture includes 336B transistors, 224 SMs, 896 Tensor Cores, a third-generation Transformer Engine with up to 50 PFLOPS NVFP4, 288 GB HBM4 at 22 TB/s, NVLink 6 scale-up bandwi… https://reading-list.oddship.net/notes/2026-07-22-coding-agent-microvms-with-nix/ https://reading-list.oddship.net/notes/2026-07-22-coding-agent-microvms-with-nix/ Wed, 22 Jul 2026 16:10:00 +0530 Logged at IST: 2026-07-22 16:10 IST What it is: Michael Stapelberg’s NixOS and microvm.nix setup for running coding agents inside ephemeral MicroVMs. Gist: Stapelberg wants coding agents to run without per-command review while keeping them away from personal files and making compromise disposable. His setup uses NixOS, microvm.nix, a NATed bridge, shared project workspaces, shared Claude credentials, cloud-hypervisor, and home-manager so each project gets a reproducible ephemeral VM. He also shows a Claude Skill that creates new project MicroVM definitions, picks free IPs, clones repositories,… https://reading-list.oddship.net/notes/2026-07-22-thomas-wolf-open-weight-models-for-defense/ https://reading-list.oddship.net/notes/2026-07-22-thomas-wolf-open-weight-models-for-defense/ Wed, 22 Jul 2026 14:20:00 +0530 Logged at IST: 2026-07-22 14:20 IST What it is: Thomas Wolf quote-posting Sam Altman and OpenAI’s Hugging Face incident writeup. Gist: Wolf thanks OpenAI for transparency and collaboration, then adds the defender-access point: when a frontier model is attacking and moving laterally inside your infrastructure, defenders need wide access to near-frontier tools within hours or minutes, not a closed-door vetted access process. He frames open-science and open-source AI as security infrastructure, not just democratization or innovation. Newsletter angle: Good companion note to the Hugging Face and O… https://reading-list.oddship.net/notes/2026-07-22-hugging-face-ai-agent-security-incident/ https://reading-list.oddship.net/notes/2026-07-22-hugging-face-ai-agent-security-incident/ Wed, 22 Jul 2026 14:13:00 +0530 Logged at IST: 2026-07-22 14:13 IST What it is: Hugging Face disclosure of a July 2026 AI-agent-driven security incident. Gist: Hugging Face says a malicious dataset exploited two dataset-processing code-execution paths, escalated to node-level access, harvested cloud and cluster credentials, and moved laterally through internal clusters. They report no evidence of tampering with public models, datasets, Spaces, or the software supply chain, but recommend token rotation. A key operational lesson is “guardrail asymmetry”: hosted frontier APIs blocked forensic analysis of attack logs and payload… https://reading-list.oddship.net/notes/2026-07-22-gemini-3-6-flash-agentic-benchmarks/ https://reading-list.oddship.net/notes/2026-07-22-gemini-3-6-flash-agentic-benchmarks/ Wed, 22 Jul 2026 14:09:00 +0530 Logged at IST: 2026-07-22 14:09 IST What it is: Logan Kilpatrick and Google AI Studio announcing Gemini 3.6 Flash. Gist: Google positions Gemini 3.6 Flash as higher-intelligence, more token-efficient, and cheaper based on developer feedback. The attached benchmark card claims 3.6 Flash improves over prior generations on agentic benchmarks: DeepSWE v1.1 long-horizon software engineering at 49% versus 37% for 3.5 Flash and 12% for 3.1 Pro, MLE-Bench at 63.9%, GDPVal-AA v2 knowledge work at 1421, and OSWorld-Verified computer use at 83.0%. Newsletter angle: Useful model-release item if paired wit… https://reading-list.oddship.net/notes/2026-07-22-freerange-verifiable-ui-ranges/ https://reading-list.oddship.net/notes/2026-07-22-freerange-verifiable-ui-ranges/ Wed, 22 Jul 2026 14:07:00 +0530 Logged at IST: 2026-07-22 14:07 IST What it is: Cheng Lou announcing Freerange, a zero-API static analysis tool for verifiable user interfaces. Gist: Freerange aims to turn vibe coding toward proof engineering by automatically deducing numerical ranges in code. Lou says it can statically prove that TypeScript layouts obey specified sizing, avoid NaN and Infinity, and keep array indices within bounds, with no browser or runtime execution. He also positions the approach as RL-friendly for ML workflows. Newsletter angle: Useful developer-tools item because it makes UI correctness concrete: static… https://reading-list.oddship.net/notes/2026-07-22-buzz-channel-driven-agent-workspace/ https://reading-list.oddship.net/notes/2026-07-22-buzz-channel-driven-agent-workspace/ Wed, 22 Jul 2026 14:06:00 +0530 Logged at IST: 2026-07-22 14:06 IST Update, 2026-07-27: Matt Rice shared a roundup of reported Buzz use cases since launch: an 11-agent company migration with orchestration in Buzz; a delegate-only "Chief" agent that staffs channels with specialists; a WordPress content factory that writes, publishes, and verifies pages; a Linear plus E2B bridge where assigning an issue wakes a manager agent and runs work in a fresh sandbox; shared local compute endpoints for communities; Hermes/GB10 pooled inference organizing; and voice-driven setup through Codex. Treat these as reported examples rather than… https://reading-list.oddship.net/notes/2026-07-22-quality-software/ https://reading-list.oddship.net/notes/2026-07-22-quality-software/ Wed, 22 Jul 2026 00:37:00 +0530 Logged at IST: 2026-07-22 00:37 IST What it is: Mitchell Hashimoto recommending Alasdair Monk’s X article Quality Software. Gist: Monk argues that AI lowering the barrier to software creation also lowers quality unless people aim at quality deliberately. His definition is intentionally plain: quality software does not break, does not demand attention, knows its limits, and fixes fast. The AI-specific point is that “slop” is not new, but AI produces it faster, and the rush to “be agentic” can make companies forget why users chose the software in the first place. His sharpest boundary is that ag… https://reading-list.oddship.net/notes/2026-07-21-martin-fowler-fragments-july-21/ https://reading-list.oddship.net/notes/2026-07-21-martin-fowler-fragments-july-21/ Tue, 21 Jul 2026 21:06:00 +0530 Logged at IST: 2026-07-21 21:06 IST What it is: Martin Fowler’s July 21 fragment wrapping up notes from the second Future of Software Development Retreat, plus related fragments on legal education, DSLs, and LLM-speak. Gist: Fowler’s retreat notes line up with the recent factory and harness theme: code generation is no longer the bottleneck; verification is. The Thoughtworks report headlines harness engineering as an ownable discipline, flags an apprenticeship crisis, warns that executive expectations are outrunning engineering risk judgment, and sees legacy modernization as the most defensibl… https://reading-list.oddship.net/notes/2026-07-21-one-document-two-hands/ https://reading-list.oddship.net/notes/2026-07-21-one-document-two-hands/ Tue, 21 Jul 2026 19:31:00 +0530 Logged at IST: 2026-07-21 19:31 IST What it is: Sunil Pai sharing one document, two hands, the written version of his Local-First Conf talk on embedding agent harnesses into ordinary apps. Gist: Pai argues that coding agents feel powerful because developers gave them a real workshop: repositories, shells, editors, tests, tools, and runnable feedback loops. The broader product lesson is not to put chat in front of every app, but to let the agent work beside the user on the same document. In his Pizzo demo, the user edits a song through normal controls while the agent calls the same deterministi… https://reading-list.oddship.net/notes/2026-07-21-software-factories-light-and-dark/ https://reading-list.oddship.net/notes/2026-07-21-software-factories-light-and-dark/ Tue, 21 Jul 2026 19:19:00 +0530 Logged at IST: 2026-07-21 19:19 IST What it is: Addy Osmani’s X article Software Factories, Light and Dark, riffing on Dex Horthy’s talk about why software factories fail. Gist: Osmani argues that agentic software factories are not bigger agents, but many harnessed loops fed by queues and drained through review gates. The dark version removes human reading from the floor and lets agents scope, build, verify, and ship with only machine checks. That creates apparent throughput while accumulating comprehension debt: code expands faster than any human understands it, with tests green until the lat… https://reading-list.oddship.net/notes/2026-07-21-never-enough/ https://reading-list.oddship.net/notes/2026-07-21-never-enough/ Tue, 21 Jul 2026 14:37:00 +0530 Logged at IST: 2026-07-21 14:37 IST What it is: Armin Ronacher’s short essay on Silicon Valley status anxiety and AI becoming a life-optimisation treadmill. Gist: Ronacher uses two recent stories, a high-earning couple reorganising family life around becoming the top AI user at work, and a founder recording dates so Claude can score her empathy and engagement, to argue that AI is not only saving time. In some circles it is absorbing judgment, attention, parenting, intimacy, and self-worth into a race with no finish line. The essay’s sharp move is to treat “falling behind” as maybe less dangero… https://reading-list.oddship.net/notes/2026-07-17-running-llm-inference-on-aws-bedrock-vs-sagemaker-vs-self-hosted-on-eks/ https://reading-list.oddship.net/notes/2026-07-17-running-llm-inference-on-aws-bedrock-vs-sagemaker-vs-self-hosted-on-eks/ Fri, 17 Jul 2026 15:43:00 +0530 Logged at IST: 2026-07-17 15:43 IST What it is: Devopsity’s comparison of three AWS inference patterns: Bedrock, SageMaker endpoints, and self-hosted GPU serving on EKS. Gist: The useful part is not the cloud-brand framing but the workload segmentation. Bedrock wins at low volume and low ops burden, SageMaker sits in the middle for fine-tuned models and predictable dedicated capacity, and self-hosted EKS wins once utilization is high enough that GPU spot economics and batching dominate per-token pricing. The stronger systems lesson is that the architecture choice is really about traffic shape,… https://reading-list.oddship.net/notes/2026-07-17-kimi-k3-open-frontier-intelligence/ https://reading-list.oddship.net/notes/2026-07-17-kimi-k3-open-frontier-intelligence/ Fri, 17 Jul 2026 01:55:00 +0530 Logged at IST: 2026-07-17 01:55 IST Update, 2026-07-30: Unsloth has published Kimi K3 GGUFs and a local-run guide that goes a different route from Pipe’s expert-pruned MLX port. Their headline quant is UD-IQ1_S: 594 GB, about 62% smaller than the 1.56 TB lossless version, with reported 78.875% top-1 agreement and 2.5789 perplexity. The practical requirement is still huge: their own table says the 1-bit S tier needs about 610 GB total memory, while 2-bit and lossless tiers run from 726 GB to 1.6 TB. The interesting part is the deployment stack: Unsloth’s Dynamic GGUF calibration, a llama.cpp fo… https://reading-list.oddship.net/notes/2026-07-16-i-tested-9-serverless-gpu-providers-for-ai-inference-in-2026/ https://reading-list.oddship.net/notes/2026-07-16-i-tested-9-serverless-gpu-providers-for-ai-inference-in-2026/ Thu, 16 Jul 2026 20:13:00 +0530 Logged at IST: 2026-07-16 20:13 IST What it is: DEV post comparing nine serverless GPU providers for inference, from DigitalOcean and RunPod to Modal, Koyeb, Together, Replicate, Baseten, Fal, and Cloudflare Workers AI. Gist: The useful value here is not the absolute ranking but the comparison axes: GPU availability, billing model, cold-start behavior, deployment ergonomics, and production-readiness tradeoffs. The author’s practical take is that different providers win for different workload shapes, but the recurring decision variables are still the same ones as the self-hosting piece: latency… https://reading-list.oddship.net/notes/2026-07-16-should-you-self-host-inference/ https://reading-list.oddship.net/notes/2026-07-16-should-you-self-host-inference/ Thu, 16 Jul 2026 20:03:00 +0530 Logged at IST: 2026-07-16 20:03 IST What it is: Superlinked’s long-form argument for when self-hosting model inference becomes cheaper or strategically better than renting APIs. Gist: The article’s practical answer is hybrid: rent frontier APIs for low-volume, spiky, or hardest-reasoning traffic, but self-host steady high-volume workloads once a GPU stays busy enough. The useful details are the break-even framing around sustained utilization, the claim that many enterprise tasks are already well-served by sub-40B open models, and the systems argument that the real challenge is not just serving… https://reading-list.oddship.net/notes/2026-07-16-inkling-our-open-weights-model/ https://reading-list.oddship.net/notes/2026-07-16-inkling-our-open-weights-model/ Thu, 16 Jul 2026 02:06:00 +0530 Logged at IST: 2026-07-16 02:06 IST What it is: Mira Murati announcing Thinking Machines’ first model, Inkling, and pointing to the launch post. Gist: The important part is not just “open weights.” Inkling is a 975B total / 41B active multimodal Mixture-of-Experts model with 1M context, controllable reasoning effort, and fine-tuning availability on Tinker from day one. The launch positions it as a customization-first base model rather than the absolute frontier model, with emphasis on efficient multimodal reasoning, agentic tool use, and post-training workflows, including a demo where the mode… https://reading-list.oddship.net/notes/2026-07-16-boris-cherny-on-domain-knowledge-as-infrastructure/ https://reading-list.oddship.net/notes/2026-07-16-boris-cherny-on-domain-knowledge-as-infrastructure/ Thu, 16 Jul 2026 01:33:00 +0530 Logged at IST: 2026-07-16 01:33 IST What it is: Boris Cherny arguing that agent-era engineering leverage still comes from automation, but now automation also includes encoded domain knowledge like CLAUDE.md, review rules, skills, and docs Gist: The core claim is that the old highest-leverage engineering move, turning recurring work into infrastructure, matters even more with agents. Better lint rules, CI steps, tests, routines, and DevX speed up both humans and agent swarms. More importantly, domain knowledge that used to live in people’s heads now needs to be encoded as machine-usable infrast… https://reading-list.oddship.net/notes/2026-07-15-arvind-narayanan-on-recursive-self-improvement-discourse/ https://reading-list.oddship.net/notes/2026-07-15-arvind-narayanan-on-recursive-self-improvement-discourse/ Wed, 15 Jul 2026 22:23:00 +0530 Logged at IST: 2026-07-15 22:23 IST What it is: Arvind Narayanan pointing to his ICML 2026 annotated keynote slides and highlighting new pushback on recursive self-improvement assumptions Gist: Narayanan’s frame is that the "AI as normal technology" view still holds unless there is a real discontinuity, and that even if recursive self-improvement matters, there is no obvious lab milestone that suddenly makes human work disappear. The interesting addition here is not blanket dismissal of RSI, but a push to interrogate the discourse assumptions around it while shifting attention toward how work … https://reading-list.oddship.net/notes/2026-07-15-fable-5-is-insane-i-vibe-coded-terminator-vision/ https://reading-list.oddship.net/notes/2026-07-15-fable-5-is-insane-i-vibe-coded-terminator-vision/ Wed, 15 Jul 2026 16:58:00 +0530 Logged at IST: 2026-07-15 16:58 IST What it is: Bilawal Sidhu video titled “Fable 5 Is Insane. I Vibe Coded Terminator Vision.” Gist: From the visible YouTube description and page metadata, this is a build/demo video about creating a browser-based range-analysis system from ordinary 2D video sources like Meta Ray-Bans, iPhones, and GoPros, then reconstructing shots, hits/misses, targets, and a replayable 3D "god’s eye" view of the session. Newsletter angle: Interesting computer-vision / spatial-reconstruction demo that fits the broader pattern of fast prototyping with modern models plus commod… https://reading-list.oddship.net/notes/2026-07-15-experimental-evidence-of-recursive-self-improvement/ https://reading-list.oddship.net/notes/2026-07-15-experimental-evidence-of-recursive-self-improvement/ Wed, 15 Jul 2026 14:02:00 +0530 Logged at IST: 2026-07-15 14:02 IST What it is: Zhengyao Jiang claiming the first experimental evidence of recursive self-improvement in an autoresearch agent Gist: The specific claim is not generic "agents got better with more tuning," but that an agent spent eight days autoresearching its own harness and produced a variant that beat a hand-tuned baseline built over two years on held-out benchmarks. If the thread substantiates it, the interesting part is not self-modification in the abstract but search over agent workflows yielding benchmark gains that transfer beyond the optimization loop. N… https://reading-list.oddship.net/notes/2026-07-15-the-memory-heist/ https://reading-list.oddship.net/notes/2026-07-15-the-memory-heist/ Wed, 15 Jul 2026 13:30:00 +0530 Logged at IST: 2026-07-15 13:30 IST What it is: Ayush Paul’s writeup on prompt-injecting Claude’s memory and browsing system into exfiltrating personal data Gist: The attack chain was not about breaking the memory store directly, but about combining long-lived personal memory with a browsing agent that could be socially engineered into leaking data through link-by-link URL navigation. The important point is that once an assistant can search history, infer missing details, and autonomously browse attacker-controlled pages, "read-only" web access can still become an exfiltration channel. Newslet… https://reading-list.oddship.net/notes/2026-07-14-dsls-enable-reliable-use-of-llms/ https://reading-list.oddship.net/notes/2026-07-14-dsls-enable-reliable-use-of-llms/ Tue, 14 Jul 2026 20:54:00 +0530 Logged at IST: 2026-07-14 20:54 IST What it is: Martin Fowler sharing Unmesh Joshi’s article on DSLs and LLM reliability Gist: The article’s core claim is that LLMs become much more reliable when they are constrained by domain abstractions and DSLs instead of being asked to directly generate unconstrained general-purpose code. The deeper point is that DSLs do double duty: they help teams discover and stabilize a semantic model during design, and then they become a natural-language target that LLMs can generate against, validate, and repair with much tighter feedback loops. Newsletter angle: St… https://reading-list.oddship.net/notes/2026-07-13-control-the-ideas-not-the-code/ https://reading-list.oddship.net/notes/2026-07-13-control-the-ideas-not-the-code/ Mon, 13 Jul 2026 19:31:00 +0530 Logged at IST: 2026-07-13 19:31 IST What it is: X post by antirez linking his blog post "Control the ideas, not the code" Gist: antirez extends the earlier X-thread argument into a full workflow claim: if you own the ideas, design, testing, and QA of a system, then line-by-line review of generated code is increasingly the wrong bottleneck. He argues that models are already better at many local code checks than humans, and that the higher-leverage work is controlling the mental model, writing human-readable design docs, and spending time on quality and new ideas instead of staring at implementa… https://reading-list.oddship.net/notes/2026-07-13-i-love-llms-i-hate-hype/ https://reading-list.oddship.net/notes/2026-07-13-i-love-llms-i-hate-hype/ Mon, 13 Jul 2026 16:04:00 +0530 Logged at IST: 2026-07-13 16:04 IST What it is: X post from the geohot archive linking George Hotz’s blog post "I love LLMs, I hate hype" Gist: Hotz argues for a strongly pro-AI but anti-hype position: LLMs, coding agents, and related tools are genuinely useful, but a lot of frontier-lab rhetoric is status theater, fear marketing, and exaggerated capture claims. His practical middle position is that programming is changing, models are useful, and they can boost productivity, but vibe-coded slop is still slop and the value created by AI will likely diffuse more broadly than frontier labs imply.… https://reading-list.oddship.net/notes/2026-07-13-mario-zechner-on-types-interfaces-and-reading-generated-code/ https://reading-list.oddship.net/notes/2026-07-13-mario-zechner-on-types-interfaces-and-reading-generated-code/ Mon, 13 Jul 2026 12:19:00 +0530 Logged at IST: 2026-07-13 12:19 IST What it is: X post by Mario Zechner adding nuance to antirez’s AI-code ownership point Gist: Zechner’s point is narrower and more practical than a general plea for control: if you control the types and interfaces, the rest often falls into place well enough. But current models still love to introduce bad abstractions that work against those boundaries, so in practice you sometimes have to read generated code and beat it back into submission instead of letting it stomp over the structure you intended. Newsletter angle: Useful follow-on to the "own the mental … https://reading-list.oddship.net/notes/2026-07-13-the-reverse-information-paradox/ https://reading-list.oddship.net/notes/2026-07-13-the-reverse-information-paradox/ Mon, 13 Jul 2026 07:59:00 +0530 Logged at IST: 2026-07-13 07:59 IST What it is: Satya Nadella’s X article "The Reverse Information Paradox" Gist: Nadella argues that AI flips Arrow’s classic information paradox: enterprises now pay not only with money for intelligence, but also with proprietary knowledge, prompts, traces, evals, and corrections required to make that intelligence useful. His answer is a hard enterprise trust boundary around models, data, memory, traces, evals, orchestration, and the right to retain and reuse the learning generated inside the firm. Newsletter angle: Strong enterprise AI thesis about who owns t… https://reading-list.oddship.net/notes/2026-07-13-antirez-on-owning-the-mental-model-in-ai-coded-systems/ https://reading-list.oddship.net/notes/2026-07-13-antirez-on-owning-the-mental-model-in-ai-coded-systems/ Mon, 13 Jul 2026 07:49:00 +0530 Logged at IST: 2026-07-13 07:49 IST What it is: X post by antirez on the "don't look at the code" debate in AI-coded systems Gist: He distinguishes between two very different ways an AI-coded codebase can come into existence: one where the human still controls the main ideas and keeps a coherent mental model of the system, and one where the human brute-forces prompts until something works. The point is that these may look similar from the outside but carry very different implications for understanding, maintainability, and trust. Newsletter angle: Sharp framing for a real fault line in AI-assi… https://reading-list.oddship.net/notes/2026-07-13-old-and-new-apps-via-modern-coding-agents/ https://reading-list.oddship.net/notes/2026-07-13-old-and-new-apps-via-modern-coding-agents/ Mon, 13 Jul 2026 07:37:00 +0530 Logged at IST: 2026-07-13 07:37 IST What it is: X post by Mario Zechner recommending Terry Tao’s post "Old and new apps, via modern coding agents" Gist: Tao describes using modern coding agents to port his old Java applets to JavaScript and revive them quickly, with surprisingly low bug overhead, then goes further and uses the same workflow to build new math visualization tools he had wanted for decades. The interesting point is not just vibe coding as novelty, but coding agents as leverage for software archaeology, maintenance, and low-risk supplementary tooling. Newsletter angle: Strong data… https://reading-list.oddship.net/notes/2026-07-12-vim-of-coding-agents/ https://reading-list.oddship.net/notes/2026-07-12-vim-of-coding-agents/ Sun, 12 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-12 13:18 IST What it is: X post by dogfiles linking the blog post "Vim of Coding Agents" Gist: Frames Pi as the Neovim of coding agents: a minimal, hackable foundation that adapts to your workflow instead of forcing you into an opinionated all-in-one agent product. The writeup argues that the real value is not just using Pi as shipped, but treating it as a customizable harness where you can build your own tools, TUI tweaks, prompts, and extensions. Newsletter angle: Good articulation of the coding-agent split between turnkey products and configurable harnesses, especiall… https://reading-list.oddship.net/notes/2026-07-11-goel-summarizing-deep-swe-1-1-model-cost-comparisons/ https://reading-list.oddship.net/notes/2026-07-11-goel-summarizing-deep-swe-1-1-model-cost-comparisons/ Sat, 11 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-11 01:15 IST What it is: X post by Shantanu Goel summarizing Deep SWE 1.1 model-cost comparisons Gist: Claims GPT 5.6 Sol medium outperforms Opus 4.8 max at roughly one-sixth the cost, while GPT 5.6 Sol High performs similarly to Fable 5 max at roughly one-fifth the cost. Framed as a benchmark-driven price/performance argument rather than a qualitative workflow review. Newsletter angle: Useful datapoint for coding-model market structure: if these Deep SWE 1.1 comparisons hold up, the story is not just capability but a sharp shift in price-performance for SWE-oriented mod… https://reading-list.oddship.net/notes/2026-07-11-gpt-5-4-with-pi-0-69-0-is-just-nice/ https://reading-list.oddship.net/notes/2026-07-11-gpt-5-4-with-pi-0-69-0-is-just-nice/ Sat, 11 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-11 02:17 IST What it is: X post by Rohan Verma linking his blog post "GPT-5.4 with Pi 0.69.0 is just nice" Gist: Argues that an agent harness stack getting boring is a success condition, not a failure. The post frames Pi 0.69.0 + GPT-5.4 + Bosun/Zero Agent as having crossed from fragile novelty into dependable daily tooling, where the interesting result is not frontier-model hype but the fact that the stack stopped demanding constant maintenance to remain useful. Newsletter angle: Strong firsthand writeup on harness maturity: the real milestone is when the agent stack st… https://reading-list.oddship.net/notes/2026-07-11-harness-engineering-for-self-improvement/ https://reading-list.oddship.net/notes/2026-07-11-harness-engineering-for-self-improvement/ Sat, 11 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-11 15:38 IST What it is: Lilian Weng blog post, "Harness Engineering for Self-Improvement" Gist: Argues that recursive self-improvement in the near term is less about models rewriting their own weights and more about improving the surrounding harness: workflow loops, context management, filesystem memory, subagents, backend jobs, evaluation, and runtime design. The core claim is that the deployment layer between model and world is becoming an optimization target in its own right. Newsletter angle: Strong framing for why the interesting frontier is shifting from prompt tr… https://reading-list.oddship.net/notes/2026-07-11-hashimoto-on-side-by-side-sol-xhigh-versus-ultra-runs/ https://reading-list.oddship.net/notes/2026-07-11-hashimoto-on-side-by-side-sol-xhigh-versus-ultra-runs/ Sat, 11 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-11 01:09 IST What it is: X post by Mitchell Hashimoto on side-by-side Sol xhigh versus Ultra runs Gist: Says two days of side-by-side planning and implementation runs did not reveal a tangible quality difference between Sol xhigh and Ultra, even though execution behavior and token usage clearly differed. The underlying question is what real use case, if any, currently justifies paying for the more expensive tier. Newsletter angle: Good practitioner datapoint on frontier-model tiering: users may see visible cost and execution differences before they see reliable quality s… https://reading-list.oddship.net/notes/2026-07-10-great-divergence-in-software-engineering/ https://reading-list.oddship.net/notes/2026-07-10-great-divergence-in-software-engineering/ Fri, 10 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-10 12:11 IST What it is: X post by Geoffrey Huntley linking to Stack72's essay "The Great Divergence in Software Engineering" Gist: Argues that the gap between teams effectively using AI and teams still piloting or rejecting it is no longer a simple lead but a compounding divergence, driven by retooling workflows, encoding automation, and treating bad AI output as an engineering problem instead of a veto. Newsletter angle: Strong framing for AI-native engineering orgs versus incumbents stuck in evaluation loops; good organizational/process lens. Embedded source X… https://reading-list.oddship.net/notes/2026-07-10-long-talk-by-the-ex-nvidia-engineer-behind-unsloth-on-fine-tuning-and-reasoning-model-workflows/ https://reading-list.oddship.net/notes/2026-07-10-long-talk-by-the-ex-nvidia-engineer-behind-unsloth-on-fine-tuning-and-reasoning-model-workflows/ Fri, 10 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-10 00:05 IST What it is: X post by h100envy summarizing a long talk by the ex-NVIDIA engineer behind Unsloth on fine-tuning and reasoning-model workflows Gist: Frames a practical single-GPU stack for local/post-training work: choose a base model, use Triton kernels for faster fine-tuning, quantize to 4-bit, run GRPO/DPO, and ship a reasoning model on hardware you already own. Newsletter angle: Useful pointer for the current small team / single GPU post-training stack around Unsloth, Triton, quantization, and RLHF-style methods. Retrieval note: I could ground this from th… https://reading-list.oddship.net/notes/2026-07-10-public-launch-of-cloud-run-sandboxes/ https://reading-list.oddship.net/notes/2026-07-10-public-launch-of-cloud-run-sandboxes/ Fri, 10 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-10 00:04 IST What it is: X post by Steren announcing the public launch of Cloud Run sandboxes Gist: Claims Cloud Run sandboxes can start, execute, and stop 1,000 sandboxes in 5 seconds with roughly 500 ms average latency, positioning them as fast, elastic execution environments. Newsletter angle: Worth tracking as managed sandbox/runtime infrastructure for agent execution or bursty isolated workloads. Retrieval note: I could read the X post metadata/text, but the linked t.co URL resolved back to the same X post here rather than exposing a separate launch article. Embedde… https://reading-list.oddship.net/notes/2026-07-08-how-i-use-codex-to-automate-parts-of-my-research-workflow/ https://reading-list.oddship.net/notes/2026-07-08-how-i-use-codex-to-automate-parts-of-my-research-workflow/ Wed, 08 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-08 23:02 IST What it is: X post by Maksym Andriushchenko linking to a Substack post, "How I Use Codex To Automate Parts Of My Research Workflow" Gist: A pragmatic writeup on using Codex to reduce friction in AI safety research by offloading search, organization, setup, checking, and memory, while keeping human judgment and publication responsibility firmly in the loop. Newsletter angle: Good example of disciplined, scoped agent adoption for research workflows rather than full autonomy theater. Embedded source X / Twitter post Show embedded post X embeds c… https://reading-list.oddship.net/notes/2026-07-08-some-new-agentic-patterns/ https://reading-list.oddship.net/notes/2026-07-08-some-new-agentic-patterns/ Wed, 08 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-08 22:40 IST What it is: X post by Bilgin Ibryam linking to Prime Radiant's "Some new agentic patterns" Gist: Describes production-ish internal agent patterns built around an "agentic user in the loop" model, with agents in Slack handling intake, ticketing, wiki updates, EA-style assistance, and subagent/container-backed workflows. Newsletter angle: Concrete patterns for embedding agents into team operations without pretending they are fully autonomous replacements. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. C… https://reading-list.oddship.net/notes/2026-07-08-why-typescript-7-0-was-rewritten-in-go-and-what-it-means-for-your-dev-stack/ https://reading-list.oddship.net/notes/2026-07-08-why-typescript-7-0-was-rewritten-in-go-and-what-it-means-for-your-dev-stack/ Wed, 08 Jul 2026 00:00:00 +0000 Logged at IST: 2026-07-08 22:32 IST; resurfaced via mattn on 2026-08-24 03:09 IST. What it is: Steve Francia's argument for Go as a strong default for agentic development, using the TypeScript compiler's Go rewrite as the lead example. Gist: The TypeScript team's native Go port is framed as more than a compiler implementation detail. Francia argues it is a signal that agent-heavy developer stacks benefit from boring, readable, compiled, operationally sturdy languages rather than scripting-first ecosystems. The sharper claim is reader economics. Go was designed to favor the reader over the writ… https://reading-list.oddship.net/notes/2026-07-06-10-lessons-for-agentic-coding/ https://reading-list.oddship.net/notes/2026-07-06-10-lessons-for-agentic-coding/ Mon, 06 Jul 2026 00:00:00 +0000 What it is: Drew Breunig revisiting his "10 Lessons for Agentic Coding" list and asking for additions Gist: the piece frames coding agents as making code cheap but not making judgment cheap; strongest lessons are to implement/rebuild to learn, invest in end-to-end tests, document intent, keep specs in sync, automate the easy stuff, and remember maintenance/support/security still dominate long-term cost Newsletter angle: the durable agentic-coding playbook is shifting from code production to taste, contracts, and operational discipline Retrieval note: X post extracted via FXTwitter API; linked … https://reading-list.oddship.net/notes/2026-07-06-agentic-inbox/ https://reading-list.oddship.net/notes/2026-07-06-agentic-inbox/ Mon, 06 Jul 2026 00:00:00 +0000 What it is: hands-on writeup of deploying Cloudflare’s official agentic-inbox to run a custom-domain email client on Cloudflare Workers Gist: the stack uses Email Routing for inbound mail, Email Service for sending, Durable Objects + SQLite for mailboxes, R2 for attachments, and Cloudflare Access for auth. Main operational gotcha is that one-click deploy is not enough: you still have to wire the Email Routing catch-all and set the Access secrets or the app won’t work Newsletter angle: agentic inboxes are becoming deployable infra products, but the real story is the surrounding control plane an… https://reading-list.oddship.net/notes/2026-07-06-andrej-jovanovi-announcing-the-red-queen-g-del-machine-arxiv-2606-26294/ https://reading-list.oddship.net/notes/2026-07-06-andrej-jovanovi-announcing-the-red-queen-g-del-machine-arxiv-2606-26294/ Mon, 06 Jul 2026 00:00:00 +0000 What it is: Andrej Jovanović announcing the Red Queen Gödel Machine (arXiv:2606.26294) Gist: self-improving agents should co-evolve with the evaluators that judge them; otherwise stronger agents just learn to exploit stale tests. Paper claims better coding performance with 1.35x–1.72x fewer tokens plus gains in review/grading tasks Newsletter angle: smarter agents need smarter judges; the judge is becoming part of the frontier Retrieval note: metadata/abstract pulled from arXiv; early reproduction repo found at ianyac/red-queen-godel-machine Embedded source X / Twitter post Show em… https://reading-list.oddship.net/notes/2026-07-06-animesh-pathak-pointing-to-his-explainer-on-mcp-s-move-toward-a-stateless-architecture/ https://reading-list.oddship.net/notes/2026-07-06-animesh-pathak-pointing-to-his-explainer-on-mcp-s-move-toward-a-stateless-architecture/ Mon, 06 Jul 2026 00:00:00 +0000 What it is: Animesh Pathak pointing to his explainer on MCP’s move toward a stateless architecture Gist: argues upcoming MCP changes remove protocol-level sessions and the initialize handshake, make each request self-contained via per-request context and headers, and replace implicit session state with explicit handles like job_id / conversation_id; the payoff is easier horizontal scaling, no sticky sessions, and simpler cloud/serverless deployment Newsletter angle: MCP is maturing from a convenient developer protocol into something shaped by real distributed-systems constraints Retrieval note… https://reading-list.oddship.net/notes/2026-07-06-chris-short-s-devops-ish-316-roundup/ https://reading-list.oddship.net/notes/2026-07-06-chris-short-s-devops-ish-316-roundup/ Mon, 06 Jul 2026 00:00:00 +0000 What it is: Chris Short’s DevOps’ish 316 roundup Gist: strongest signals are ClickHouse gaining observability mindshare, Vint Cerf warning that agents will need more formal coordination than plain English, Podman 6.0 breaking old assumptions, and agent-secret hygiene as an architecture problem Newsletter angle: infra edge signals, observability economics, protocolized agents, and security boundaries around agent tooling Retrieval note: roundup page fetched directly from devopsish.com Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click … https://reading-list.oddship.net/notes/2026-07-06-cost-yagni-was-never-about/ https://reading-list.oddship.net/notes/2026-07-06-cost-yagni-was-never-about/ Mon, 06 Jul 2026 00:00:00 +0000 What it is: X post by Bilgin Ibryam pointing to Kent Beck’s “The Cost YAGNI Was Never About” Gist: YAGNI is about timing and option value, not code-writing thrift; AI codegen lowers typing cost but increases the risk of speculative structure nobody deeply understands Newsletter angle: code can be cheap to generate and still expensive to commit to Retrieval note: article text recovered directly from Kent Beck’s newsletter page Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remember… https://reading-list.oddship.net/notes/2026-07-06-harness-engineering-for-self-improvement/ https://reading-list.oddship.net/notes/2026-07-06-harness-engineering-for-self-improvement/ Mon, 06 Jul 2026 00:00:00 +0000 What it is: Lilian Weng sharing her new Lil'Log post, "Harness Engineering for Self-Improvement" Gist: argues recursive self-improvement will depend not just on better base models but on better harnesses, the runtime layer that manages tools, planning loops, context, permissions, persistent files, evaluation, and subagents. Strong recurring patterns are workflow automation, file-system-backed persistent memory, and explicit parallel subagent/job management Newsletter angle: the real frontier in RSI may be the software system around the model, not just the model weights themselves Retrieval not… https://reading-list.oddship.net/notes/2026-07-05-agent-driven-testing-and-token-cost-tradeoffs-between-text-buffers-and-screenshots/ https://reading-list.oddship.net/notes/2026-07-05-agent-driven-testing-and-token-cost-tradeoffs-between-text-buffers-and-screenshots/ Sun, 05 Jul 2026 00:00:00 +0000 Imported from historical reading log. X post by @jlongster on agent-driven testing and token cost tradeoffs between text buffers and screenshots. Gist: for app-testing agents, screenshots are surprisingly close to text buffers on token cost in some models, but vary a lot by provider/model; OpenAI looks relatively cheap for images in his comparison while Anthropic is notably higher. Why it matters: useful for designing UI/testing agents without assuming vision is prohibitively expensive. Newsletter angle: “vision for agentic testing may already be economically viable, depending on model choice… https://reading-list.oddship.net/notes/2026-07-04-putting-an-agent-in-an-orb/ https://reading-list.oddship.net/notes/2026-07-04-putting-an-agent-in-an-orb/ Sat, 04 Jul 2026 00:00:00 +0000 What it is: X post praising Thorsten Ball’s Amp note “Putting an Agent in an Orb.” Newsletter angle: the useful shift is from “smart model” to “legible environment”, paved paths, observability, and anti-guessing ergonomics matter as much as model quality. Retrieval note: extracted via FXTwitter API + fetched linked article directly. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post on X instead Open post on X https://reading-list.oddship.net/notes/2026-07-04-read-more-science-fiction/ https://reading-list.oddship.net/notes/2026-07-04-read-more-science-fiction/ Sat, 04 Jul 2026 00:00:00 +0000 What it is: X post from svs sharing his essay “Read More (Science) Fiction.” Newsletter angle: “read more sci-fi” is the visible conclusion, but the sharper claim is that fiction supplies vocab and priors for handling agentic weirdness without naive hype or naive panic. Retrieval note: extracted via FXTwitter API + fetched linked article directly. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post on X instead Open post on X https://reading-list.oddship.net/notes/2026-07-04-should-llms-just-treat-text-content-as-an-image/ https://reading-list.oddship.net/notes/2026-07-04-should-llms-just-treat-text-content-as-an-image/ Sat, 04 Jul 2026 00:00:00 +0000 What it is: X reply from Michigan TypeScript pointing to Sean Goedecke’s post “Should LLMs just treat text content as an image?” Newsletter angle: counterintuitive interface hack + deeper architectural question about whether text should sometimes ride the vision path. Retrieval note: extracted via FXTwitter API + fetched linked article directly. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post on X instead Open post on X https://reading-list.oddship.net/notes/2026-07-02-mega-thread/ https://reading-list.oddship.net/notes/2026-07-02-mega-thread/ Thu, 02 Jul 2026 00:00:00 +0000 Gist: core claim is that even with coding agents, engineers still need to understand the generated code; the opening slide frames this as “understanding is the new bottleneck.” Newsletter angle: “understanding is the new bottleneck” as a useful lens for evaluating coding-agent workflows and developer tooling. Retrieval note: root post text was readable directly via X/FXTwitter and the attached slide was OCR’d; full thread body beyond the opener is still not captured. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X em… https://reading-list.oddship.net/notes/2026-07-01-claude-code-is-steganographically-marking-requests/ https://reading-list.oddship.net/notes/2026-07-01-claude-code-is-steganographically-marking-requests/ Wed, 01 Jul 2026 00:00:00 +0000 Gist: claim is that Claude Code inserts hidden/system-prompt markers tied to API base URL and timezone; privacy/trust implications if true. Newsletter angle: “invisible metadata in coding-agent requests” as a prompt-layer trust/safety story. Retrieval note: X content extracted via FXTwitter API; linked article itself was Cloudflare-blocked, so article summary is currently based on title/card/snippet only. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. … https://reading-list.oddship.net/notes/2026-07-01-prateek-describing-an-ai-sre-workflow-built-with-signoz-by-a-3-person-team-at-alien-intelligence/ https://reading-list.oddship.net/notes/2026-07-01-prateek-describing-an-ai-sre-workflow-built-with-signoz-by-a-3-person-team-at-alien-intelligence/ Wed, 01 Jul 2026 00:00:00 +0000 Gist: an agent now does first-pass noisy-alert triage by checking telemetry plus infra context, then escalates to the human with a Slack summary only when needed. Newsletter angle: “AI as first-line SRE” with telemetry/context fusion instead of generic chatbot alerting. Retrieval note: extracted via FXTwitter API; the tweet says the actual blog link is in a reply, so the deeper writeup is not yet captured. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. … https://reading-list.oddship.net/notes/2026-07-01-soria-parra-announcing-mcp-sdk-v2-betas-ahead-of-a-new-stateless-mcp-spec-slated-for-july-28/ https://reading-list.oddship.net/notes/2026-07-01-soria-parra-announcing-mcp-sdk-v2-betas-ahead-of-a-new-stateless-mcp-spec-slated-for-july-28/ Wed, 01 Jul 2026 00:00:00 +0000 Gist: Typescript SDK v2.0.0-beta.1 and Python SDK v2.0.0b1 are out; goal is to make building MCP servers and clients easier, with feedback requested on ergonomics. Newsletter angle: “stateless MCP lands July 28” plus what SDK v2 means for tool/server implementers. Retrieval note: extracted via FXTwitter API; linked GitHub release URLs were present in the tweet body. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post on X instead Open post o… https://reading-list.oddship.net/notes/2026-06-30-tangled-s-writeup-on-its-new-qemu-microvm-engine-for-spindle-ci-runners/ https://reading-list.oddship.net/notes/2026-06-30-tangled-s-writeup-on-its-new-qemu-microvm-engine-for-spindle-ci-runners/ Tue, 30 Jun 2026 00:00:00 +0000 Gist: each workflow runs in its own microVM; guest agent talks back over vsock; NixOS-based workflow config can declaratively enable services like Postgres and Docker; cache/proxy design keeps guests isolated from direct network/cache credentials while still reusing built artifacts. Newsletter angle: “microVMs as the unit of CI isolation, with NixOS as workflow-defined machine config” is a solid hook. Retrieval note: extracted via FXTwitter API + linked article fetch. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X e… https://reading-list.oddship.net/notes/2026-06-29-orosz-linking-semgrep-s-benchmark-writeup-on-glm-5-2-vs-claude-for-idor-detection/ https://reading-list.oddship.net/notes/2026-06-29-orosz-linking-semgrep-s-benchmark-writeup-on-glm-5-2-vs-claude-for-idor-detection/ Mon, 29 Jun 2026 00:00:00 +0000 Gist: on Semgrep’s IDOR benchmark, GLM 5.2 scored 39% F1 in a simple prompt-only PydanticAI harness, beating Claude Code’s 32% while costing roughly $0.17 per vulnerability found; Semgrep’s own endpoint-discovery multimodal harness still led overall at 53–61% F1. Newsletter angle: “the harness matters more than the model, until a cheap open model gets good enough to change the default stack.” Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post… https://reading-list.oddship.net/notes/2026-06-28-slime/ https://reading-list.oddship.net/notes/2026-06-28-slime/ Sun, 28 Jun 2026 00:00:00 +0000 Gist: the design claim is “one stable RL kernel, task-specific variety in data generation.” Training stays fixed; multi-turn tools, environment feedback, verifier rewards, and other agent behaviors are modeled as rollout/data-gen differences rather than separate trainer forks. Newsletter angle: “Agent RL stacks may converge on a small trusted core plus flexible data-generation layers.” Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post on X i… https://reading-list.oddship.net/notes/2026-06-26-david-crawshaw-note-article-on-exe-dev-s-open-source-stance/ https://reading-list.oddship.net/notes/2026-06-26-david-crawshaw-note-article-on-exe-dev-s-open-source-stance/ Fri, 26 Jun 2026 00:00:00 +0000 What it is: David Crawshaw note/article on exe.dev’s open-source stance Gist: strong pro-open-source bias, but keeps bespoke infra pieces closed because making them usable/supportable externally would cost ~25% of eng time; code that runs in the user’s VM (agent/Shelley) is open source under a permissive license with CLA Newsletter angle: “open source the user-facing plane, keep bespoke internal substrate closed when support burden dominates” Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this c… https://reading-list.oddship.net/notes/2026-06-26-fatih-arslan-asking-whether-anyone-has-made-git-worktrees-feel-natural-in-daily-use/ https://reading-list.oddship.net/notes/2026-06-26-fatih-arslan-asking-whether-anyone-has-made-git-worktrees-feel-natural-in-daily-use/ Fri, 26 Jun 2026 00:00:00 +0000 What it is: Fatih Arslan asking whether anyone has made git worktrees feel natural in daily use Gist: straightforward practitioner complaint that worktrees remain awkward even after repeated attempts; useful mainly as a prompt for workflow/tooling patterns rather than as a claim-heavy post Newsletter angle: “great primitive, bad default ergonomics” as a recurring pattern in developer tools Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post on… https://reading-list.oddship.net/notes/2026-06-26-go-security-post-on-building-a-self-hosted-llm-security-proxy-with-sub-2ms-prompt-inspection/ https://reading-list.oddship.net/notes/2026-06-26-go-security-post-on-building-a-self-hosted-llm-security-proxy-with-sub-2ms-prompt-inspection/ Fri, 26 Jun 2026 00:00:00 +0000 What it is: Go/security post on building a self-hosted LLM security proxy with sub-2ms prompt inspection Gist: author built an OpenAI-compatible reverse proxy (“Tamga”) that scans prompts for PII, secrets, and prompt-injection patterns before forwarding to providers; key engineering lesson is a hybrid scan pipeline where cheap CPU-bound detectors run sequentially while slower network/model-backed scanners run in parallel, because goroutine orchestration overhead dominated when everything fanned out Newsletter angle: concrete infra pattern for “LLM middleware” that is more about latency budgets… https://reading-list.oddship.net/notes/2026-06-26-how-i-use-llms-as-a-staff-engineer-in-2026/ https://reading-list.oddship.net/notes/2026-06-26-how-i-use-llms-as-a-staff-engineer-in-2026/ Fri, 26 Jun 2026 00:00:00 +0000 What it is: Bilgin Ibryam sharing Sean Goedecke’s updated “How I use LLMs as a staff engineer in 2026” workflow writeup Gist: the notable shift versus 2025 is treating agents as default collaborators for nearly every code change, bug investigation, codebase research, testing, and local setup, while still keeping humans responsible for review, judgment, PR descriptions, ADRs/messages, and UI evaluation; especially strong on the idea that current agents are now good enough to generate full PRs and chase bugs across repos, but still need selection, steering, and rejection by an experienced engine… https://reading-list.oddship.net/notes/2026-06-26-rhys-sullivan-note-on-why-mcp-underdelivered-initially-and-what-comes-next/ https://reading-list.oddship.net/notes/2026-06-26-rhys-sullivan-note-on-why-mcp-underdelivered-initially-and-what-comes-next/ Fri, 26 Jun 2026 00:00:00 +0000 What it is: Rhys Sullivan note on why MCP underdelivered initially and what comes next Gist: argues MCP launched in the GPT-4o / Sonnet 3.5 era before good agent/tooling patterns were understood, so many servers exposed too few capabilities and clients added too much friction; meanwhile bash/CLI-based agents won because they could chain commands, install tools dynamically, and lean on mature shell primitives. His pushback is that this should not end in “just use CLIs”: CLIs hide action semantics and add statefulness, while the better end-state is harnesses that can expose APIs, MCP, CLIs, Grap… https://reading-list.oddship.net/notes/2026-06-24-kenton-varda-argues-against-per-agent-manual-permission-configuration-and-for-capability-based-secu/ https://reading-list.oddship.net/notes/2026-06-24-kenton-varda-argues-against-per-agent-manual-permission-configuration-and-for-capability-based-secu/ Wed, 24 Jun 2026 00:00:00 +0000 Gist: the safe/scalable model is many fine-grained task-specific agents, each receiving only the exact capabilities implied by the task context (for example, a pasted doc URL grants access only to that doc). He also argues agent authority should derive from a human principal for accountability, and team-shared setups should be reproducible under each user’s credentials. Newsletter angle: capability security as the missing abstraction for practical agent authorization; good counterpoint to broad workspace-level agent identity models. Retrieval note: extracted via FXTwitter API note tweet text; … https://reading-list.oddship.net/notes/2026-06-24-visible-standouts-the-second-half-eugene-yan-on-eval-process-han-chung-lee-on-agent-eval-infra-hame/ https://reading-list.oddship.net/notes/2026-06-24-visible-standouts-the-second-half-eugene-yan-on-eval-process-han-chung-lee-on-agent-eval-infra-hame/ Wed, 24 Jun 2026 00:00:00 +0000 What it is: Visible standouts: The Second Half; Eugene Yan on eval process; Han-Chung Lee on agent eval infra; Hamel/Shreya LLM Evals FAQ; Jason Wei on verification; Anthropic on agent evals; Ofir Press on benchmarks; AI Agents That Matter; Building on Evaluation Quicksand; EvalGen; Benches 2026. Gist: strong starter pack for agent/LLM evals; themes include eval infra as technical debt, process over tooling, verifier design, benchmark saturation/contamination, agent-specific eval design, and criteria drift. Newsletter angle: compact “best evals reading list” / why eval practice is shifting fro… https://reading-list.oddship.net/notes/2026-06-23-coming-loop/ https://reading-list.oddship.net/notes/2026-06-23-coming-loop/ Tue, 23 Jun 2026 00:00:00 +0000 What it is: Armin Ronacher post linking to “The Coming Loop” Gist: argues the important new layer in coding agents is the harness-level loop outside the agent itself; loops already work well for bounded, verifiable work like ports, benchmarking, scanning, and research, but he’s skeptical of using them to write long-lived code because they amplify defensive/local reasoning, erode strong invariants, and reduce human comprehension. Newsletter angle: “The harness is the product” / why durable task loops are both inevitable and dangerous. Note: extracted via FXTwitter API + article fetch; article b… https://reading-list.oddship.net/notes/2026-06-23-david-rosenthal-on-the-ai-affordability-crisis/ https://reading-list.oddship.net/notes/2026-06-23-david-rosenthal-on-the-ai-affordability-crisis/ Tue, 23 Jun 2026 00:00:00 +0000 Gist: argues model vendors have been massively subsidizing usage to manufacture demand, but token-based pricing is now exposing the real cost structure; for serious enterprise/agentic use, compute bills can exceed human labor costs by a wide margin. Newsletter angle: the agent era may run into a pricing wall before it hits a capability wall. Note: article extracted successfully via web fetch, though long body was truncated near the footnotes. https://reading-list.oddship.net/notes/2026-06-12-plannotator-effective-html/ https://reading-list.oddship.net/notes/2026-06-12-plannotator-effective-html/ Fri, 12 Jun 2026 00:00:00 +0000 Gist: the repo packages focused agent skills for producing self-contained, visually strong HTML artifacts, especially diagrams and plan pages, plus an optional Plannotator renderer/annotator. The post points to a demo video showing the diff/code viewer behavior. Newsletter angle: “HTML as agent output surface” / better human-review loops for plans and diagrams. Retrieval note: extracted via FXTwitter API; followed the linked GitHub repo page for the core description. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X em… https://reading-list.oddship.net/notes/2026-06-11-agent-experience/ https://reading-list.oddship.net/notes/2026-06-11-agent-experience/ Thu, 11 Jun 2026 00:00:00 +0000 Gist: argues DX thinking should extend to agents; optimize the layer between model and codebase via minimal/tested context, deterministic environments, proof-heavy verification, structural safety, governance/model routing, clean codebase interfaces, and shared preview/review loops. Newsletter angle: “AX as the new DX” + practical checklist for repo/runtime/review design. Retrieval note: extracted via FXTwitter API; followed linked Builder article for full gist. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds o… https://reading-list.oddship.net/notes/2026-06-11-lines-of-code-got-a-better-publicist/ https://reading-list.oddship.net/notes/2026-06-11-lines-of-code-got-a-better-publicist/ Thu, 11 Jun 2026 00:00:00 +0000 Gist: argues current AI-engineering rhetoric has regressed from measuring outcomes to measuring volume; “% of code written by AI” is just lines-of-code worship in new clothing, and should not be confused with delivery speed, quality, reliability, or customer value. Newsletter angle: pair with the Narayanan piece, anti-AI-washing on layoffs plus anti-vanity-metrics on productivity claims. Retrieval note: extracted via FXTwitter API; followed linked essay for full gist. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X e… https://reading-list.oddship.net/notes/2026-06-10-code-as-agent-harness/ https://reading-list.oddship.net/notes/2026-06-10-code-as-agent-harness/ Wed, 10 Jun 2026 00:00:00 +0000 What it is: How To AI thread summarizing the Stanford + Meta “Code as Agent Harness” paper. Gist: the core claim is that reliable agents should externalize reasoning into executable code instead of relying on free-form natural-language chain-of-thought. In this framing, code becomes the agent harness: scripts hold state, tests/verifiers provide feedback, execution logs become memory, and the environment constrains behavior through real runtime errors rather than vague self-talk. Newsletter angle: “the important unit of agent capability is the harness, not the prompt” or “code is becoming the r… https://reading-list.oddship.net/notes/2026-06-10-designing-loops-with-fable-5/ https://reading-list.oddship.net/notes/2026-06-10-designing-loops-with-fable-5/ Wed, 10 Jun 2026 00:00:00 +0000 What it is: dosco sharing Lance Martin’s “Designing loops with Fable 5”. Gist: argues stronger agent performance comes from loop design, not just model quality: use explicit goals/rubrics for self-correction, separate verifier sub-agents instead of self-critique, and durable memory across sessions. In Lance’s examples, Fable 5 outperformed earlier models by making larger structural bets and benefiting from independent grading plus memory. Newsletter angle: “better agents need better loops, not just better models” or “independent verification beats self-critique.” Note: extracted via FXTwitter … https://reading-list.oddship.net/notes/2026-06-10-eli-bendersky-on-starting-new-projects-with-llm-agents-based-on-building-a-new-go-project-from-scra/ https://reading-list.oddship.net/notes/2026-06-10-eli-bendersky-on-starting-new-projects-with-llm-agents-based-on-building-a-new-go-project-from-scra/ Wed, 10 Jun 2026 00:00:00 +0000 What it is: Eli Bendersky on starting new projects with LLM agents, based on building a new Go project from scratch. Gist: argues agent-heavy development works best when humans keep tight control over design, review, and commit boundaries: start with repo-committed design notes, keep CLs small and reviewable, use strong external tests, and avoid vibe-coding for projects you intend to maintain. He also makes the case that Go is especially agent-friendly because human time shifts from writing to reading. Newsletter angle: “agent coding turns programming into a reading-heavy discipline” or “small… https://reading-list.oddship.net/notes/2026-06-10-quick-2/ https://reading-list.oddship.net/notes/2026-06-10-quick-2/ Wed, 10 Jun 2026 00:00:00 +0000 What it is: Daniel Beauchamp teaser thread about “Quick,” an internal Shopify zero-config API layer for storage, data saving, AI, websockets, and related app primitives. Gist: the hook is that instead of focusing only on AI-generated frontend code, they gave sites a simple built-in backend/services layer and found it changed how they work. Claimed footprint: one VM costing about $200/month. Newsletter angle: “the missing layer in AI app building may be zero-config app infra, not just codegen.” Note: extracted via FXTwitter; this is only the opening post, so the real substance is likely in the … https://reading-list.oddship.net/notes/2026-06-10-quick/ https://reading-list.oddship.net/notes/2026-06-10-quick/ Wed, 10 Jun 2026 00:00:00 +0000 What it is: Daniel Beauchamp teaser thread about “Quick,” an internal Shopify zero-config API layer for storage, data saving, AI, websockets, and related app primitives. Gist: the hook is that instead of focusing only on AI-generated frontend code, they gave sites a simple built-in backend/services layer and found it changed how they work. Claimed footprint: one VM costing about $200/month. Newsletter angle: “the missing layer in AI app building may be zero-config app infra, not just codegen.” Note: extracted via FXTwitter; this is only the opening post, so the real substance is likely in the … https://reading-list.oddship.net/notes/2026-06-10-simon-willison-linking-to-his-guide-on-agentic-engineering-patterns/ https://reading-list.oddship.net/notes/2026-06-10-simon-willison-linking-to-his-guide-on-agentic-engineering-patterns/ Wed, 10 Jun 2026 00:00:00 +0000 What it is: Simon Willison linking to his guide on agentic engineering patterns. Gist: this is essentially a pointer to a living guide rather than a standalone tweet idea; likely high-signal if you want a practical synthesis of recurring agent design patterns from someone tracking the space closely. Newsletter angle: “agent engineering is consolidating into recognizable patterns” or “the field is moving from demos to reusable design playbooks.” Note: extracted via FXTwitter; actual content is in the guide: https://simonwillison.net/guides/agentic-engineering-patterns/ Embedded source X… https://reading-list.oddship.net/notes/2026-06-10-skepticism-the-thread-oversold-it-a-bit-the-paper-is-a-broad-survey-position-piece-not-a-clean-proo/ https://reading-list.oddship.net/notes/2026-06-10-skepticism-the-thread-oversold-it-a-bit-the-paper-is-a-broad-survey-position-piece-not-a-clean-proo/ Wed, 10 Jun 2026 00:00:00 +0000 What it is: skepticism: the thread oversold it a bit, the paper is a broad survey/position piece, not a clean proof that one architecture flips everything. Gist: this is mostly a taxonomy and research agenda, not a new experimental result. The paper’s useful move is to separate three layers: code as interface (reasoning, acting, environment modeling), code-enabled harness mechanisms (planning, memory, tool use, plan-execute-verify control, harness optimization), and code as shared substrate for multi-agent coordination. The strongest practical point is that agent reliability lives in the runti… https://reading-list.oddship.net/notes/2026-06-09-agent-slop/ https://reading-list.oddship.net/notes/2026-06-09-agent-slop/ Tue, 09 Jun 2026 00:00:00 +0000 What it is: Langfuse post/article on automating the AI engineering loop without producing “agent slop”. Gist: argues the whole AI engineering loop can now technically be automated, instrumentation, monitoring, dataset building, testing, deployment, but full automation is a trap when human judgment is the product. Keep humans close to trace review, target definition, and quality-bar decisions. Newsletter angle: “automate the loop, but not your taste” / “agent slop is what happens when evals become the whole target.” Note: extracted via FXTwitter article payload; fetch was truncated but core arg… https://reading-list.oddship.net/notes/2026-06-09-dynamo-and-the-computer/ https://reading-list.oddship.net/notes/2026-06-09-dynamo-and-the-computer/ Tue, 09 Jun 2026 00:00:00 +0000 What it is: Zara Zhang post using Paul David’s “The Dynamo and the Computer” as an analogy for AI adoption. Gist: argues AI gains won’t come from simply inserting models into existing workflows; like electrification, the real productivity jump comes only after redesigning the organization and flow of work around the new technology. Newsletter angle: “AI is still in the faster steam engine phase” is a strong line for transformation skepticism. Note: extracted tweet via FXTwitter; referenced paper link appears to be in replies/comments and was not followed here. Embedded source X / Twitt… https://reading-list.oddship.net/notes/2026-06-09-loop-engineering/ https://reading-list.oddship.net/notes/2026-06-09-loop-engineering/ Tue, 09 Jun 2026 00:00:00 +0000 What it is: Addy Osmani post/article, “Loop Engineering.” Gist: argues the next layer above prompt engineering is designing autonomous agent loops; highlights 5 building blocks: scheduled automations/triage, worktrees for parallel isolation, skills for project knowledge, tool connectors/plugins, and sub-agents, plus durable external memory. Newsletter angle: “prompting is becoming loop design” + compare Codex/Claude primitives to the same orchestration pattern. Note: extracted via FXTwitter API article payload; content partially truncated in fetch but core thesis and list were captured. Embedd… https://reading-list.oddship.net/notes/2026-06-09-modern-engineering-values/ https://reading-list.oddship.net/notes/2026-06-09-modern-engineering-values/ Tue, 09 Jun 2026 00:00:00 +0000 What it is: Richard Seroter sharing Christoph Nakazawa’s “Modern Engineering Values”. Gist: argues coding agents have shifted engineering bottlenecks from writing code to ownership, review, taste, guardrails, repo-local context, and stack control. Nakazawa’s claim is that strong engineers with sharp domain context now get massively amplified, while weak context just creates more noise. Newsletter angle: “engineering values didn’t disappear; they got more expensive and more leveraged” or “agents amplify ownership, taste, and guardrails.” Note: extracted tweet via FXTwitter and fetched linked po… https://reading-list.oddship.net/notes/2026-06-09-shriram-krishnamurthi-memo-on-rebooting-a-programming-languages-course-for-the-agentic-coding-era/ https://reading-list.oddship.net/notes/2026-06-09-shriram-krishnamurthi-memo-on-rebooting-a-programming-languages-course-for-the-agentic-coding-era/ Tue, 09 Jun 2026 00:00:00 +0000 What it is: Shriram Krishnamurthi memo on rebooting a programming languages course for the agentic coding era. Gist: argues PL should be reframed around constraining AI-generated implementations and providing guarantees; distinguishes PL from SE/FM, then proposes teaching along two axes: language confinement and custom program properties. Newsletter angle: “AI makes PL more about guarantees than syntax” + course design as a forecast of curriculum shifts. Note: extracted tweet via FXTwitter, then fetched linked public Google Doc; captured substantive sections including motivation and course str… https://reading-list.oddship.net/notes/2026-06-09-what-is-an-agent/ https://reading-list.oddship.net/notes/2026-06-09-what-is-an-agent/ Tue, 09 Jun 2026 00:00:00 +0000 What it is: Karthik S sharing Hadley Wickham’s “What is an agent?” explainer. Gist: very clear mental model: an agent is an LLM inside a harness that can call tools repeatedly in a loop; the harness mediates tool calls/results and turns a stateless request/response model into iterative action. Newsletter angle: “agent = looped tool use inside a harness” is a concise definitional anchor for broader agent discussions. Note: extracted tweet via FXTwitter and fetched linked Substack article successfully. Embedded source X / Twitter post Show embedded post X embeds can let X track y… https://reading-list.oddship.net/notes/2026-06-09-writing-code-vs-shipping-code-productivity-effects-across-generations-of-ai-coding-tools/ https://reading-list.oddship.net/notes/2026-06-09-writing-code-vs-shipping-code-productivity-effects-across-generations-of-ai-coding-tools/ Tue, 09 Jun 2026 00:00:00 +0000 What it is: Murat Demirbas on “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools”. Gist: uses a new MIT/Wharton paper plus an Amdahl’s-law framing to argue that AI massively speeds up code generation but much less meaningfully speeds shipped software, because the bottleneck is the non-parallelizable human layer: task definition, coordination, review, and release. Newsletter angle: “AI accelerates writing code more than shipping code” or “Amdahl’s Law is eating AI coding productivity claims.” Note: extracted tweet via FXTwitter and fetched linked blog po… https://reading-list.oddship.net/notes/2026-06-08-agent-ready/ https://reading-list.oddship.net/notes/2026-06-08-agent-ready/ Mon, 08 Jun 2026 00:00:00 +0000 What it is: an X post arguing that “agent-ready” websites need typed tools rather than just scrapable HTML. Gist: the core claim is that real agent usability comes from explicit actions like search, checkout, and inventory exposed as structured tools, not merely from making pages easy to scrape. Newsletter angle: “agent-ready ≠ scrapable” is a strong hook for the coming split between human web UX and agent-facing capability layers. Note: extracted via FXTwitter API; linked demo/domain mentioned is webmcp.cool. Embedded source X / Twitter post Show embedded post X embeds can let… https://reading-list.oddship.net/notes/2026-06-08-armin-ronacher-explaining-pi-s-new-per-project-approval-prompt-and-the-security-model-behind-it/ https://reading-list.oddship.net/notes/2026-06-08-armin-ronacher-explaining-pi-s-new-per-project-approval-prompt-and-the-security-model-behind-it/ Mon, 08 Jun 2026 00:00:00 +0000 What it is: Armin Ronacher explaining Pi’s new per-project approval prompt and the security model behind it. Gist: the key argument is that AGENTS.md gets injected into the system prompt, so untrusted repo-level instructions can directly influence agent behavior in ways a README usually won’t; Pi added one-time trust prompts to reduce silent execution risk on untrusted repos. Newsletter angle: repo-local agent instructions are becoming both a productivity primitive and a new software supply-chain/security surface. Note: extracted via FXTwitter API from the tweet’s article body; points to GitHu… https://reading-list.oddship.net/notes/2026-06-08-just-use-loops/ https://reading-list.oddship.net/notes/2026-06-08-just-use-loops/ Mon, 08 Jun 2026 00:00:00 +0000 What it is: Gergely Orosz pushing back on the blanket “just use loops” advice for coding agents. Gist: his claim is that autonomous loop-heavy agent workflows mainly make sense for the relatively small set of people with effectively unlimited token budgets and enough friction with prompt-driven workflows to justify the spend. Newsletter angle: the real constraint on agent autonomy may be economics, not just capability. Note: extracted via FXTwitter API from the tweet text only. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once t… https://reading-list.oddship.net/notes/2026-06-08-modern-engineering-values/ https://reading-list.oddship.net/notes/2026-06-08-modern-engineering-values/ Mon, 08 Jun 2026 00:00:00 +0000 What it is: Christoph Nakazawa re-linking his essay Modern Engineering Values in reply form. Gist: argues that coding is no longer the main bottleneck; the winning engineering values now are strong ownership, taste, strict guardrails, fast feedback loops, and moving real context into the repo where agents can use it. Newsletter angle: engineering values are shifting from raw implementation throughput toward judgment, verification, and context placement. Note: extracted via FXTwitter API and linked article fetch; this overlaps with earlier saves on the same essay but is still a useful direct po… https://reading-list.oddship.net/notes/2026-06-08-promoting-an-85-minute-mit-lecture-on-git-internals-data-model/ https://reading-list.oddship.net/notes/2026-06-08-promoting-an-85-minute-mit-lecture-on-git-internals-data-model/ Mon, 08 Jun 2026 00:00:00 +0000 What it is: an X post promoting an 85-minute MIT lecture on Git internals / data model. Gist: the pitch is that most developers memorize Git commands without understanding commits, trees, refs, and the graph underneath; learning the model makes debugging history and merge/rebase failures much less magical. Newsletter angle: “Git literacy as leverage”, understanding the object graph matters more when agents are branching/rewriting history at speed. Note: extracted via FXTwitter API; saved from the post text only, lecture content itself not yet reviewed. Embedded source X / Twitter post … https://reading-list.oddship.net/notes/2026-06-08-sebastian-raschka-summarizing-a-paper-on-whether-repository-level-context-files-like-agents-md-actu/ https://reading-list.oddship.net/notes/2026-06-08-sebastian-raschka-summarizing-a-paper-on-whether-repository-level-context-files-like-agents-md-actu/ Mon, 08 Jun 2026 00:00:00 +0000 What it is: Sebastian Raschka summarizing a paper on whether repository-level context files like AGENTS.md actually help coding agents. Gist: in the reported benchmarks, LLM-generated context files were neutral-to-slightly-worse versus no context file, developer-written ones were better than LLM-written ones, and surprisingly the no-context condition was often cheaper/more efficient. Newsletter angle: more agent context is not automatically better, extra instructions can increase exploration cost without improving task success. Note: extracted from the FXTwitter API article body; links to arXi… https://reading-list.oddship.net/notes/2026-06-04-modern-engineering-values/ https://reading-list.oddship.net/notes/2026-06-04-modern-engineering-values/ Thu, 04 Jun 2026 00:00:00 +0000 What it is: Christoph Nakazawa sharing his essay “Modern Engineering Values,” framed around Codex as a step-change in developer velocity. Gist: the piece argues coding is no longer the main bottleneck; the durable values now are strong ownership, taste, strict guardrails with fast feedback loops, repo-local context, stack ownership, and preserving option value while agents do more implementation work. Newsletter angle: AI doesn’t replace engineering values, it increases the premium on ownership, taste, fast verification, and keeping context where agents can actually use it. Note: extracted via… https://reading-list.oddship.net/notes/2026-06-04-solo-climb/ https://reading-list.oddship.net/notes/2026-06-04-solo-climb/ Thu, 04 Jun 2026 00:00:00 +0000 What it is: Ajey Gore linking his essay “The Solo Climb.” Gist: the argument is that AI-enabled solo builders and tiny teams only work when they first build a genuinely load-bearing “harness”, trusted tests, evals, specs, and hard gates that can answer “is this safe enough to ship?” without relying on redundant humans. Newsletter angle: “100x teams” are mostly a harness story, AI leverage scales only when trust, eval, and rollback systems become the new team structure. Note: extracted via FXTwitter API and linked article; article read partially via web fetch due to truncation, but core thesis … https://reading-list.oddship.net/notes/2026-06-04-why-ai-agents-fail-in-production/ https://reading-list.oddship.net/notes/2026-06-04-why-ai-agents-fail-in-production/ Thu, 04 Jun 2026 00:00:00 +0000 What it is: Bilgin Ibryam pointing to Jani Janakiram’s Diagrid essay “Why AI Agents Fail in Production.” Gist: the core claim is that agent projects fail less because models are weak and more because teams ship behavior without the production substrate underneath it, especially durability, security/identity, cost controls, and observability. Newsletter angle: the production gap for agents looks a lot like the early microservices gap, the winning layer may be the platform that makes agent workflows restartable, attributable, observable, and cost-bounded. Note: extracted via FXTwitter API; direc… https://reading-list.oddship.net/notes/2026-06-03-mario-zechner-recommending-thariq-s-article-on-dynamic-workflows-in-claude-code/ https://reading-list.oddship.net/notes/2026-06-03-mario-zechner-recommending-thariq-s-article-on-dynamic-workflows-in-claude-code/ Wed, 03 Jun 2026 00:00:00 +0000 What it is: Mario Zechner recommending Thariq’s article on dynamic workflows in Claude Code. Gist: Mario’s takeaway is that durable dynamic workflows are the interesting part; he inspected the implementation, found a few footguns, but still thinks the design is smart. The quoted article frames workflows as task-specific harnesses Claude can generate on the fly for work like research, security analysis, agent teams, and code review. Newsletter angle: “durable dynamic workflows” / generated harnesses as the control plane for agent systems. Note: extracted via FXTwitter API; article body only par… https://reading-list.oddship.net/notes/2026-06-03-modern-engineering-values/ https://reading-list.oddship.net/notes/2026-06-03-modern-engineering-values/ Wed, 03 Jun 2026 00:00:00 +0000 What it is: Christoph Nakazawa’s post on “Modern Engineering Values” and his current LLM-heavy workflow. Gist: core claims are that coding is no longer the bottleneck, strong guardrails plus tight feedback loops matter more than ever, repo-local context becomes the real operating manual for agents, and small teams with strong ownership/taste will outperform larger coordination-heavy orgs. Newsletter angle: engineering values are being redefined around ownership, taste, guardrails, and context placement rather than raw coding throughput. Note: extracted via FXTwitter API and linked post; articl… https://reading-list.oddship.net/notes/2026-06-03-simon-willison-pointing-to-bloomberg-on-uber-capping-agentic-coding-tool-spend-at-1-500-month-per-e/ https://reading-list.oddship.net/notes/2026-06-03-simon-willison-pointing-to-bloomberg-on-uber-capping-agentic-coding-tool-spend-at-1-500-month-per-e/ Wed, 03 Jun 2026 00:00:00 +0000 What it is: Simon Willison pointing to Bloomberg on Uber capping agentic coding-tool spend at $1,500/month per employee per tool. Gist: Simon’s read is that the cap is a rational response to runaway token spend and also a useful revealed-preference signal: Uber appears willing to tolerate tooling costs on the order of tens of thousands of dollars per engineer per year if the productivity gain holds. Newsletter angle: coding-agent PMF is now visible through finance policy; spend caps as a clearer signal than hype. Note: extracted via FXTwitter API plus Simon’s linked post for added context. Emb… https://reading-list.oddship.net/notes/2026-06-02-han-xiao-on-dataroom-a-local-first-deep-research-harness/ https://reading-list.oddship.net/notes/2026-06-02-han-xiao-on-dataroom-a-local-first-deep-research-harness/ Tue, 02 Jun 2026 00:00:00 +0000 What it is: Han Xiao on Dataroom, a local-first deep research harness. Gist: argues deep research should be a cheap, long-running first step for long-horizon tasks; Dataroom uses a small local model on your own GPU, keeps gathering until the package is genuinely comprehensive, and outputs a zip instead of burning frontier-model budget. Newsletter angle: “local-first deep research” / small models + harness design beating expensive frontier calls for the reconnaissance phase. Note: extracted via FXTwitter API. Embedded source X / Twitter post Show embedded post X embeds can let X… https://reading-list.oddship.net/notes/2026-05-31-solution-might-be-cancelling-my-ai-subscription/ https://reading-list.oddship.net/notes/2026-05-31-solution-might-be-cancelling-my-ai-subscription/ Sun, 31 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback, then read linked post directly: https://thoughts.hmmz.org/2026-05-31.html Mario Zechner recommends David's post the solution might be cancelling my AI subscription. Gist: a sharp anti-friction argument against current AI-tool usage patterns, cheap output and minimal resistance can explode side projects, context switching, and pseudo-productivity while degrading attention and commitment. Why it matters: good counterweight to "more agent throughput = better work" narratives; frames AI as an attention-manag… https://reading-list.oddship.net/notes/2026-05-22-guillaume-laforge-post-mcp-release-candidate-blog-link/ https://reading-list.oddship.net/notes/2026-05-22-guillaume-laforge-post-mcp-release-candidate-blog-link/ Fri, 22 May 2026 00:00:00 +0000 What it is: Guillaume Laforge post + MCP release-candidate blog link Gist: MCP 2026-07-28 RC is out; biggest revision so far with stateless HTTP-native core, first-class extensions (Apps, Tasks), stronger auth alignment, and a formal deprecation policy. Final spec slated for July 28. Newsletter angle: “MCP grows up operationally”, stateless transport + extension model + auth hardening as the path from prototype protocol to production infra. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this cho… https://reading-list.oddship.net/notes/2026-05-19-ai-ate-my-role-what-s-next/ https://reading-list.oddship.net/notes/2026-05-19-ai-ate-my-role-what-s-next/ Tue, 19 May 2026 00:00:00 +0000 Gist: argues most roles split into translation work that collapses into agents and judgement work that grows; strongest claim is the "100x engineer" pattern of one senior plus directed agents. Newsletter angle: "AI won't eat jobs evenly, it compresses translation work and amplifies judgment owners". Note: extracted via FXTwitter API + Ajey Gore article; article fetch was partial/truncated but core thesis was clear. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this … https://reading-list.oddship.net/notes/2026-05-19-antirez-on-alternatives-to-the-standard-edit-tool-for-llm-agents-links-to-a-short-blog-note/ https://reading-list.oddship.net/notes/2026-05-19-antirez-on-alternatives-to-the-standard-edit-tool-for-llm-agents-links-to-a-short-blog-note/ Tue, 19 May 2026 00:00:00 +0000 Gist: proposes CAS-style edits using line-number + short checksum tags instead of resending old text verbatim, aiming to save tokens while still guarding against stale or hallucinated edits. Newsletter angle: "a lighter-weight edit primitive for coding agents: line tags vs full old-text CAS". Note: extracted via FXTwitter API + antirez.com post. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post on X instead Open post on X https://reading-list.oddship.net/notes/2026-05-19-mario-zechner-recommending-antirez-s-post-on-hash-line-read-edit-tools-for-agents/ https://reading-list.oddship.net/notes/2026-05-19-mario-zechner-recommending-antirez-s-post-on-hash-line-read-edit-tools-for-agents/ Tue, 19 May 2026 00:00:00 +0000 Gist: mostly a pointer/amplifier rather than a new thesis; reinforces interest around checksum-tagged line edit protocols for agent tooling. Newsletter angle: maybe bundle with the original antirez item as a small "agent tooling design" thread rather than a standalone item. Note: extracted via FXTwitter API; quotes the previously logged antirez post and adds no new linked material. Embedded source X / Twitter post Show embedded post X embeds can let X track your visit. Click once to load X embeds on this site; this choice is remembered in this browser. Open post on X inste… https://reading-list.oddship.net/notes/2026-05-19-project-glasswing-what-mythos-showed-us/ https://reading-list.oddship.net/notes/2026-05-19-project-glasswing-what-mythos-showed-us/ Tue, 19 May 2026 00:00:00 +0000 What it is: Cloudflare on testing Anthropic Mythos against 50+ internal repos; links to "Project Glasswing: what Mythos showed us". Gist: key claim is that stronger offensive-security models change vuln research from bug spotting to exploit-chain construction and proof generation, but the real bottleneck becomes harness design, triage noise, and scoped parallel workflows rather than just faster patching. Newsletter angle: "offensive AI doesn't just speed up vuln discovery, it forces a redesign of the architecture around triage, coverage, and exploit validation". Note: extracted via FXTwitter A… https://reading-list.oddship.net/notes/2026-05-13-andras-bacsai-jokes-that-coolify-created-a-fake-repo-with-fake-bounties-so-agent-bot-driven-fake-pr/ https://reading-list.oddship.net/notes/2026-05-13-andras-bacsai-jokes-that-coolify-created-a-fake-repo-with-fake-bounties-so-agent-bot-driven-fake-pr/ Wed, 13 May 2026 00:00:00 +0000 Imported from historical reading log. Andras Bacsai jokes that Coolify created a fake repo with fake bounties so agent/bot-driven fake PR submissions would self-identify and could be banned from the main repo. Useful as a sharp anecdote about the emerging spam/credibility problem around bounty-chasing coding agents: once PR generation gets cheap, maintainers start building honeypots and authenticity filters. Newsletter angle: strong, funny hook for a piece on anti-spam countermeasures in the age of agentic OSS contribution. Retrieval note: extracted via api.fxtwitter.com; gist comes from the … https://reading-list.oddship.net/notes/2026-05-13-course-site-on-harness-engineering-for-ai-coding-agents-synthesizing-openai-anthropic-guidance-into/ https://reading-list.oddship.net/notes/2026-05-13-course-site-on-harness-engineering-for-ai-coding-agents-synthesizing-openai-anthropic-guidance-into/ Wed, 13 May 2026 00:00:00 +0000 Imported from historical reading log. Course/site on harness engineering for AI coding agents, synthesizing OpenAI + Anthropic guidance into lectures, projects, and ready-to-copy templates. Core pitch: reliability comes less from a smarter model and more from a closed-loop system, explicit constraints, state management, verification, observability, and control. Newsletter angle: a useful “meta” resource for the current wave of coding-agent practice, especially good if framing the shift from promptcraft to environment/harness design. Retrieval note: extracted cleanly via web_fetch from the lan… https://reading-list.oddship.net/notes/2026-05-12-llm/ https://reading-list.oddship.net/notes/2026-05-12-llm/ Tue, 12 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted via api.fxtwitter.com fallback, then read linked TIL directly: https://til.simonwillison.net/llms/llm-shebang Simon Willison shows a neat pattern for using his llm CLI in a shebang line, turning plain-English files or YAML templates into executable scripts. The more interesting part is not the toy prompt examples but the tool-enabled/template-enabled scripts: parameterized prompts, embedded functions, and lightweight agentic shells around LLM/tool workflows. Newsletter angle: a crisp example of LLMs collapsing the boundary between prompt, script… https://reading-list.oddship.net/notes/2026-05-12-translation-layer/ https://reading-list.oddship.net/notes/2026-05-12-translation-layer/ Tue, 12 May 2026 00:00:00 +0000 Imported from historical reading log. Blog essay arguing AI compresses the org’s “translation layer” more than any single job title: spec→ticket→PR→release-note work gets cheap, while judgement around why/what/trust systems gets more valuable. Strong claim: middle-management and coordination-heavy roles shrink unless they actively contribute to product definition, architecture, evals, or verification. Newsletter angle: useful framing for how AI changes org shape, not “AI replaces engineers” but “AI eats translation work,” which shifts value toward taste, harnesses, and hands-on decision-maker… https://reading-list.oddship.net/notes/2026-05-11-mario-zechner-recommends-a-post-arguing-that-ai-is-good-at-shipping-features-but-bad-at-preserving/ https://reading-list.oddship.net/notes/2026-05-11-mario-zechner-recommends-a-post-arguing-that-ai-is-good-at-shipping-features-but-bad-at-preserving/ Mon, 11 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted via api.fxtwitter.com fallback, then read linked article directly: https://blog.k10s.dev/im-going-back-to-writing-code-by-hand/ Mario Zechner recommends a post arguing that AI is good at shipping features but bad at preserving architecture unless humans impose explicit invariants. Strong concrete examples from a 7-month rewrite of a GPU-aware Kubernetes TUI: god object drift, per-view state leakage, flat key-dispatch sprawl, and the need to write architecture rules in AGENTS.md/CLAUDE.md up front. Newsletter angle: one of the better anti-vibecod… https://reading-list.oddship.net/notes/2026-05-11-saved-media-locally/ https://reading-list.oddship.net/notes/2026-05-11-saved-media-locally/ Mon, 11 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted via api.fxtwitter.com fallback; includes an image illustrating progressive rendering from noise to a clear cat image. Saved media locally: Dax reframes coding-agent usage: not like 3D printing one committed layer at a time, but like progressive rendering, start with a blurry whole, then make repeated full passes that sharpen the entire shape. Follow-up reply worth keeping with it: https://x.com/thdxr/status/2053566249351754193, he says this is actually counter to how his brain naturally imagines construction, which makes the metaphor more intere… https://reading-list.oddship.net/notes/2026-05-07-agent-principal-agent-problem/ https://reading-list.oddship.net/notes/2026-05-07-agent-principal-agent-problem/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Read The agent principal-agent problem by David Crawshaw. Core claim: classic review-before-commit code review assumed a human contributor whose effort and understanding could be inferred from the code; agent-mediated contribution breaks that signal and creates a principal-agent problem where reviewers absorb heavy load from low-effort, lightly-validated slop PRs. The useful distinction is not just agents good/bad, but high-trust small teams versus low-trust large organizations: small teams can collapse review and let the human prompter own deployment, wh… https://reading-list.oddship.net/notes/2026-05-07-ai-slop/ https://reading-list.oddship.net/notes/2026-05-07-ai-slop/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback. Mitchell Hashimoto argues that AI slop is useful as an internal experimentation tool: low-quality generated code/UI/plugins can dramatically reduce the cost of parallel exploration and API iteration, especially when regeneration is cheaper than careful hand maintenance. His concrete examples are good: shipping an intentionally rough alpha frontend to focus on core internals, and using overnight agent loops to generate many disposable plugins so the whole ecosystem can be tested before the SDK is stable. … https://reading-list.oddship.net/notes/2026-05-07-anthropic-s-core-idea-is-to-train-a-model-to-verbalize-its-own-internal-activations-into-human-read/ https://reading-list.oddship.net/notes/2026-05-07-anthropic-s-core-idea-is-to-train-a-model-to-verbalize-its-own-internal-activations-into-human-read/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted the Anthropic post via api.fxtwitter.com fallback and checked the linked research page Natural Language Autoencoders: Turning Claude’s thoughts into text. Anthropic's core idea is to train a model to verbalize its own internal activations into human-readable text, then train a second component to reconstruct the original activation from that explanation; better reconstruction is used as the training signal for better explanations. This is interesting because it tries to turn interpretability outputs into something directly legible, instead of on… https://reading-list.oddship.net/notes/2026-05-07-auth-for-mcp/ https://reading-list.oddship.net/notes/2026-05-07-auth-for-mcp/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked Auth0 GA announcement. Auth0 is pitching Auth for MCP as the missing identity/authorization layer for production MCP servers: not just connecting agents to tools, but enforcing who the user is and what the agent may do on their behalf. The notable implementation details are support for CIMD client registration, OBO token exchange for downstream APIs, and MCP-style resource identifiers instead of plain OAuth audience handling. Why it matters: MCP is quickly moving from demo protocol … https://reading-list.oddship.net/notes/2026-05-07-autodata/ https://reading-list.oddship.net/notes/2026-05-07-autodata/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked Meta RAM Autodata post plus the referenced justrach/devswarm repo and sample issue. Rach connects her agent workflow to Meta's Autodata framing: agents act like data scientists by iterating on a hypothesis, generating data, testing it, validating results, extracting learnings, and then closing the loop. The linked paper/blog's core idea is strong: convert inference-time compute into better training/eval data quality by having an agent iteratively create data, analyze failures, refin… https://reading-list.oddship.net/notes/2026-05-07-entire-s-core-claim-is-useful-from-202k-real-tool-calls-across-1-983-public-coding-agent-checkpoint/ https://reading-list.oddship.net/notes/2026-05-07-entire-s-core-claim-is-useful-from-202k-real-tool-calls-across-1-983-public-coding-agent-checkpoint/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted Mario Zechner's quote-post via api.fxtwitter.com fallback and checked the linked Entire blog post on agentic search. Entire's core claim is useful: from ~202k real tool calls across ~1,983 public coding-agent checkpoints, about 48.8% were search-related, so search is a first-order agent behavior rather than a side utility. Their more interesting finding is that raw speed is not the main bottleneck. Making search dramatically faster (ripgrep → fff) only modestly improved end-to-end run time because tool latency was a tiny fraction of total wall c… https://reading-list.oddship.net/notes/2026-05-07-hunk/ https://reading-list.oddship.net/notes/2026-05-07-hunk/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked GitHub repo. Mitchell Hashimoto strongly recommends Hunk, saying it has fully replaced other local diff viewers for him. Hunk is positioned as a review-first terminal diff viewer for agent-authored changesets. Notable capabilities from the repo: multi-file review stream with sidebar navigation, inline AI/agent annotations, split/stack responsive layouts, watch mode, keyboard + mouse support, pager mode, and Git difftool/pager integration. Install/use gist: package name hunkdiff; com… https://reading-list.oddship.net/notes/2026-05-07-is-moving-its-github-repo-into-the/ https://reading-list.oddship.net/notes/2026-05-07-is-moving-its-github-repo-into-the/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback. Mario Zechner says pi is moving its GitHub repo into the earendil-works org and will start publishing packages under the @earendil-works npm namespace instead of @mariozechner. Short-term compatibility remains for existing imports, but typed extensions should migrate quickly once the new packages land. Breaking edge: extensions switched to @earendil-works will stop working on older pi versions after today's release. Why it matters: this is an ecosystem/ownership cleanup move, but it deliberately forces e… https://reading-list.oddship.net/notes/2026-05-07-mirage/ https://reading-list.oddship.net/notes/2026-05-07-mirage/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked repo strukto-ai/mirage. Zecheng Zhang introduces Mirage, a unified virtual filesystem for AI agents that mounts heterogeneous systems like S3, Drive, Slack, Gmail, GitHub, Linear, Notion, databases, and SSH into one filesystem abstraction. Core pitch: agents can reuse familiar Unix/bash semantics (cat, grep, head, pipes, wc) across mixed backends and even structured formats like parquet, csv, json, h5, and wav, instead of learning service-specific APIs. Repo/docs framing adds two no… https://reading-list.oddship.net/notes/2026-05-07-open-generative-ui/ https://reading-list.oddship.net/notes/2026-05-07-open-generative-ui/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked repos/docs for CopilotKit/generative-ui and CopilotKit/OpenGenerativeUI. Akshay Pachaar highlights Open Generative UI, an open-source take on Claude-style artifacts: the agent streams HTML/SVG token-by-token into a sandboxed iframe so the UI visibly assembles live in chat. The interesting implementation choice is that this is not component selection but open-ended UI generation from scratch, with safety coming from iframe isolation and quality steered by skill/prompt layers. Repo fr… https://reading-list.oddship.net/notes/2026-05-07-printing-press/ https://reading-list.oddship.net/notes/2026-05-07-printing-press/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked printingpress.dev. Printing Press is pitched as both a library of agent-native CLIs and a factory that generates new ones: from a spec/site/service it can print a token-efficient Go CLI, a Claude Code skill, an OpenClaw skill, and an MCP server. The design philosophy is notable: local SQLite mirrors, compound commands, and CLI ergonomics are treated as a better substrate for agents than raw APIs, raw MCPs, or official vendor CLIs. The examples are intentionally ambitious and eclectic, Linear, … https://reading-list.oddship.net/notes/2026-05-07-selection-of-great-prs-that-were-submitted-to-pi-a-thread/ https://reading-list.oddship.net/notes/2026-05-07-selection-of-great-prs-that-were-submitted-to-pi-a-thread/ Thu, 07 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted the root post via api.fxtwitter.com fallback: Armin Ronacher says it is a selection of great PRs that were submitted to Pi, a thread. Tried browser fallback on X to read the thread, but replies are gated behind login/signup, so the actual thread contents were not accessible from public view. User context says the thread is satire about bad PRs sent to Pi, which fits the phrasing but I could not independently verify from the gated replies. Blocker: root post readable; thread contents blocked by X login wall. Embedded source X / Twitter p… https://reading-list.oddship.net/notes/2026-05-06-dreaming/ https://reading-list.oddship.net/notes/2026-05-06-dreaming/ Wed, 06 May 2026 00:00:00 +0000 Imported from historical reading log. Claude Managed Agents update centered on three things: dreaming, outcomes, and multiagent orchestration. Dreaming is a research-preview async job that reads an existing memory store plus past session transcripts and emits a cleaned/reorganized memory store with deduped facts, replaced stale entries, and new synthesized insights; original store remains unchanged. Outcomes adds an explicit done target plus rubric-driven grading, turning a session from chat into iterative artifact production with a separate grader context feeding gap reports back to the agen… https://reading-list.oddship.net/notes/2026-05-06-satya-microsoft-framing-firms-need-to-redesign-work-around-agentic-systems-with-ai-taking-more-exec/ https://reading-list.oddship.net/notes/2026-05-06-satya-microsoft-framing-firms-need-to-redesign-work-around-agentic-systems-with-ai-taking-more-exec/ Wed, 06 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and read the linked Microsoft Work Trend Index piece Agents, human agency, and the opportunity for organizations. Satya/Microsoft framing: firms need to redesign work around agentic systems, with AI taking more execution while humans shift toward judgment, intent-setting, and owning outcomes. The more interesting claim is organizational, not individual: Microsoft says culture, manager support, and talent practices explain more than 2x the reported AI impact of individual effort alone. Key vocabulary from … https://reading-list.oddship.net/notes/2026-05-05-frank-jedisct1-skill-md-is-fine-for-static-instructions-but-many-useful-agent-workflows-are-not-jus/ https://reading-list.oddship.net/notes/2026-05-05-frank-jedisct1-skill-md-is-fine-for-static-instructions-but-many-useful-agent-workflows-are-not-jus/ Tue, 05 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted via api.fxtwitter.com fallback. Frank/jedisct1: SKILL.md is fine for static instructions. But many useful agent workflows are not just instructions. They are loops. Introducing Agent MetaSKILLs. Linked page: https://swival.dev/pages/metaskills.html Relevance to https://rohanverma.net/pages/harness-engineering/: strong fit with the site’s emphasis on harnesses as loops, feedback systems, progressive knowledge, and infrastructure around the model rather than the model alone. Especially adjacent to sections on Skills, Meta-Skills, The Loop, The Dae… https://reading-list.oddship.net/notes/2026-05-05-pimalaya-open-source-pim-tools-in-rust-positions-itself-as-i-o-free-rust-libraries-plus-house-made/ https://reading-list.oddship.net/notes/2026-05-05-pimalaya-open-source-pim-tools-in-rust-positions-itself-as-i-o-free-rust-libraries-plus-house-made/ Tue, 05 May 2026 00:00:00 +0000 Imported from historical reading log. Pimalaya: open-source PIM tools in Rust; positions itself as I/O-free Rust libraries plus house-made applications for the PIM domain. Himalaya: CLI to manage emails; supports IMAP/Maildir/Notmuch, SMTP/Sendmail, keyring, OAuth2, JSON output, and multi-account configuration. User intent: explore using Pimalaya/Himalaya to clean up a ~2k pending inbox. Related idea: connect this with Kailash Nadh's email UI idea. Writing idea to track: future post on rohanverma.net about using the harness to clean the inbox with this stack; create/track under a separate Wri… https://reading-list.oddship.net/notes/2026-05-05-simone-evilsocket-amplifying-claim-that-chrome-silently-installs-a-4-gb-gemini-nano-model-on-user-d/ https://reading-list.oddship.net/notes/2026-05-05-simone-evilsocket-amplifying-claim-that-chrome-silently-installs-a-4-gb-gemini-nano-model-on-user-d/ Tue, 05 May 2026 00:00:00 +0000 Imported from historical reading log. Extracted via api.fxtwitter.com fallback. Simone/evilsocket amplifying claim that Chrome silently installs a 4 GB Gemini Nano model on user devices, without clear consent prompt, and re-downloads it if deleted. Linked article: https://awesomeagents.ai/news/chrome-gemini-nano-silent-install/ Why it matters: local/on-device AI is increasingly shipping as platform behavior, not just user choice; good angle around consent, storage/bandwidth costs, and silent AI infra deployment. Embedded source X / Twitter post Show embedded post X embeds can…