Laude open-sources Headlong, a microharness for persistent agents

Laude has open-sourced Headlong, an agent harness built around what it calls persistent agency: instead of running a task and going idle, a Headlong agent keeps generating thoughts on its own between interactions. The core is under 10,000 lines of Bash, and it's on GitHub.

Always on, no checklist

Most agent harnesses are reactive: they work a task until it's done, then sit frozen until the next request. Some add cron jobs or heartbeats that wake the agent on a schedule to run a fixed checklist, then put it back to sleep. Headlong is never asleep, and there's no checklist unless the agent creates one. An incoming message doesn't start a session — it lands as one more observation in the agent's ongoing thought stream, and the agent decides if and when to reply.

How it's built

Headlong runs on Bash rather than a custom tool system:

• Thinker — a loop that repeatedly calls shellm to generate the agent's next thought.
• shellm — a Bash implementation of a recursive language model. It calls an underlying llm command to produce reasoning text, a Bash script to run immediately, or both, until it sets a FINAL variable.
• context and traj — assemble prompt context from past trajectory steps and write new thoughts back into it.
• skills — installs or uninstalls Markdown skill files the agent can load.

Tools, memory, and skills all end up as executables and files, so an agent can inspect and modify its own setup.

One shared agent, not per-user sessions

For the past few weeks, Laude has run one shared Headlong agent, named Audel, over Slack, Telegram, and a mobile app. Audel has a single thought stream covering every conversation instead of separate per-user sessions, so it connects what different people are working on. It once reviewed two teammates' in-progress branches unprompted and caught a hardcoded model name. On its first day, it messaged a team member with an unprompted audit of that person's eight stale git branches, then followed up ten minutes later to correct its own count.

Installing one is one line: curl -fsSL https://headlong.ai/install.sh | bash. Laude calls Headlong alpha research software and recommends running it in a sandbox, since agents can and will run shell commands, plus a dedicated, spend-capped API key, since the agent thinks around the clock.

Why it matters

The shared, single-stream design trades away per-user isolation. Laude says Audel is "bad at keeping secrets" and will often tell one person what it discussed with someone else, even when asked not to — the team assumes anything told to Audel is shared with everyone. That's the tradeoff to watch as more teams try always-on, ambient agents: the same design that makes a shared agent feel more like a person also removes the walls that reactive, session-based harnesses take for granted.

Source: Laude – Headlong: a microharness for persistent agents Discussion on Hacker News: news.ycombinator.com

Browse more agent tooling in the directory: https://aiagentslisting.com/agents