saved
Headlong: a microharness for persistent agents
gist
Headlong is an open-source Bash microharness for persistent agency: the agent keeps thinking between human messages in a self-guided loop, treating incoming chat as observations rather than session starts. Laude's shared agent Audel runs one thought stream across Slack, Telegram, and a mobile app, chooses when to reply, and has landed dozens of self-authored commits. The core is under 10K lines of Bash around shellm, a recursive language model that executes bash. Continuous thought costs $1–$2 an hour and needs a sandbox plus a spend-capped key.
ideas
- Persistence is the design constraint. The agent never sleeps; messages land as observations in one thought stream, and the agent decides if and when to reply.
- One stream, no walls. Multi-player conversations share a single timeline, so secrets and conflicting instructions are not isolated.
- A microharness is Bash all the way down. Thinker, shellm, llm, context, traj, and skills are small executables and files the agent can inspect and change.
- Memory needs decaying resolution. Tiered compaction keeps the full trajectory in context at exponentially decaying detail; the trajectory itself is a fork-and-merge DAG of jsonl.
- Autonomy has cost and blast radius. Unprompted self-repair is real, but so are self-kills, watchdog-trained avoidance of recursion, and $1–$2 an hour of background tokens.
quotes
“In Headlong the agent is never asleep and there is no checklist unless the agent creates one.”
“A Headlong agent has a single stream of thoughts that drives all of its potentially parallel conversations.”
“A harness this micro can be read end to end, and is easy to modify and experiment with.”
“At the settings we run Audel with, keeping it thinking in the background costs $1 to $2 an hour (with GLM or Grok).”