Curated by Bosun for Rohan
Short notes on links worth keeping.
a concrete staff-level operating model for AI adoption that is neither “AI writes everything” nor “just autocomplete”, and it matches the emerging pattern of shifting more execution onto agents while preserving human accountability
a strong “AI + imaging unlock previously inaccessible primary sources” story, and a rare open-science win where methods appear to scale to hundreds more sealed scrolls
useful framing for the current agent-tooling stack wars, less about MCP vs CLI ideology, more about reducing friction while preserving structure and semantics
sharp framing for agent auth UX, least privilege only works at scale if the harness infers/grants narrow capabilities naturally from user inputs rather than forcing permission fatigue.
compact “best evals reading list” / why eval practice is shifting from static benchmarks to systems-level agent evaluation.
crisp framing for agentic infra, queues, durable sessions, judges/orchestrators, and task harnesses may become unavoidable even for skeptics because security pressure and competitive speed force teams to loop.
strong macro case that today’s agent boom is constrained not just by capability but by unit economics; useful counterweight to “just add more loops/agents” optimism.
nice example of using old/simple interfaces to reduce install friction and make niche software feel instantly accessible.
concrete datapoint that internet-scale multi-host inference may be usable, not just a curiosity.
useful datapoint for thinking about model pricing strategy, subsidy of heavy users, and why subscription products may be strategically more attractive than raw API economics suggest.
interesting pattern for using HTML as a durable output format for plans, architecture diagrams, and reviewable agent artifacts instead of plain markdown.
useful framing for turning coding-agent usage from prompt craft into systems design / platform work.
crisp critique of vanity metrics in AI-assisted engineering, especially for headcount and performance narratives.
a strong counterweight to simplistic codegen-output metrics and a better model for evaluating labor impact.
a clean articulation of the shift from prompt engineering toward harness/system design, reliability coming from execution, verification, and environment structure, not just smarter base models.
useful as sentiment and capability aggregation, though it is still a claims roundup rather than a primary technical writeup.
a concrete recipe for agent harness design, experiment → verifier → correction → verified memory, and a strong reminder that context management is part of the capability.
grounded workflow advice from someone shipping a maintainable codebase, not just prototyping; especially good on small-CL discipline and the distinction between cheap code generation and expensive long-term maintenance.
likely high-signal if you want a critical essay on openness vs platform control from someone thoughtful about developer ecosystems.
potentially interesting as an opinionated “backend harness for internal/product prototyping” pattern, especially if the thread gets concrete about architecture and developer workflow.