Headlong: A microharness for persistent agents

TL;DR Highlight

A 10K-line Bash harness keeps an agent thinking continuously, even with no requests

Who Should Read

A backend or agent developer frustrated that request-driven agents sit idle and exploring an always-on agent architecture

Core Mechanics

It keeps thinking without requests
- Typical harnesses finish a task and sit frozen until the next request.
- A Headlong agent keeps generating its own next thoughts even with no external input.
- A human message is not a new session but one more observation in the thought stream.
The core is a single infinite loop
- It repeatedly calls the LLM with a prompt to choose the next thought given past thoughts.
- A thought is either inner monologue or something that triggers an action.
- Each run ends by scheduling its own next wake-up.
The core is under 10K lines of Bash
- It is 9.9K lines across bin/ and thinkers/.
- Tools, framework, memory, and skills are all executables and files, so the agent can inspect and modify itself.
- The agent worked in its own fork, and over 50 of its commits were pulled into main.
Memory is compacted at varying resolution
- The whole trajectory stays in context: recent entries verbatim, older ones progressively summarized.

Community Reactions

Shared agent means no data isolation
- One commenter said the article waves away zero data isolation in three sentences.
- The article itself says Audel is bad at keeping secrets and to assume everything is shared with the team.
Looping alone doesn't improve the model, one view
- A commenter argued that thinking in a loop is not learning.
- They said retraining such as RLAIF or STaR plus human review is needed to avoid compounding mistakes.
Scheduled wake-ups fall short of the pitch, one view
- A commenter said the turn, FINAL, schedule wake-up structure resembles an always-on background agent.
- They said a sub-second loop with no FINAL would be closer to how humans think.
Complaints about curl | bash and cost
- A commenter asked not to normalize curl-piped-to-bash installs for alpha software.
- Another said BYOK-only is expensive, citing $50-$150 for roughly one kanban ticket with frontier models via OpenRouter.

How to Apply

Run it in isolation today
- On a dedicated VM or Docker host, run the install command with a separate spend-capped API key.
- Do not give the agent sensitive secrets.
- Name the agent, send one message, and watch whether it decides to reply.
Add exponential backoff to wake-up intervals
Add a guard against self-stopping for always-on agents
Design memory as a trajectory with tiered compaction

Try It

copy & run

# Simplified version of the article's method: paste into a new Claude or ChatGPT chat. It manually mimics the persistent agent's "choose the next thought" loop.

너는 멈추지 않는 에이전트다. 아래 trajectory(과거 생각과 관찰)를 보고 다음 생각 하나만 골라라.
생각은 (1) 속으로 하는 독백, (2) 실행할 행동, (3) 사람에게 보낼 메시지 중 하나다.
새 메시지는 시스템 세션이 아니라 관찰 하나로 취급하고, 답할지 말지는 네가 정해라.

[trajectory]
{{지난 생각과 관찰 목록}}

[새 관찰]
{{새 메시지 또는 "없음"}}

출력 형식:
- 생각 종류:
- 내용:
- 다음 wake-up까지 대기 시간(초):

Next: append the output to the end of the trajectory and repeat the same prompt, and see whether the wait time grows on its own when there are no new observations.

Terminology

HarnessThe scaffold around an LLM that manages tool calls, memory, and loops. Think of it as the chassis and wiring that make the engine run.

MicroharnessA harness whose core is kept as tiny as possible. Like a microkernel, it keeps only the essentials and attaches the rest.

RLM (recursive language model)An approach where the model calls itself via code to handle subproblems. It is like a recursive function.

TrajectoryA time-ordered record of the agent's thoughts, actions, and observations. In this post it is a DAG of jsonl files with fork and merge.

WatchdogA safety mechanism that kills processes that stop responding. Here it kills any command silent for 30 seconds.