Headlong and the always-on agent harness

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, trajectory as append-only jsonl DAGs with fork/merge, context as a projection of that trajectory, and memory/skills as files. The framing is very Unix: give the model Bash as the universal tool interface and keep the harness small enough that the agent can inspect and modify itself.

The most useful part is the operational honesty. Laude's shared agent, Audel, reportedly worked in its own fork and contributed more than 50 commits back to main, including one 48-minute autonomous diagnosis and fix of its own broken recall process. But they also call out the real costs: background thinking at roughly $1-2/hour, weak secrecy boundaries in a shared single mind, accidental service stops, and failed early self-delegation.

That makes Headlong less interesting as a one-line install and more interesting as a design probe: what does it take to operate an agent that is not a session, but a living process?

Newsletter angle: Strong agents/infra item: persistent agents move the problem from “can an agent complete this task?” to “can I safely operate an always-on, self-modifying process with spend, privacy, memory, and sandbox controls?”