Headlong: open-source persistent agents in under 10K lines of Bash
Laude Institute and MIT ship an Apache-2.0 agent that thinks between messages and runs on any LLM key for ~$1–2/hr — but shares one memory across all users.
Copy markdown
An agent that thinks between your messages
Headlong runs a continuous inner-monologue loop instead of waking only when pinged: messages arrive as observations in an ongoing thought stream, and the agent sets its own priorities and decides when to reply. Released today under Apache-2.0 by the Laude Institute and MIT, with a one-line install.
Bash is the entire tool layer
The core is under 10K lines of Bash. Its `shellm` engine implements recursive LMs — the agent reasons by writing and running shell commands, so there's no separate tool schema to maintain. History lives in a jsonl trajectory DAG with tiered summarization: nothing is dropped, older turns just compress.
One agent, many people, one timeline
A single Headlong agent can hold Slack, Telegram and web-chat conversations at once, experiencing all of them in one unified thought stream. It runs on an Anthropic, OpenAI, Gemini or OpenRouter key and costs roughly $1–2/hour thinking in the background, backing off exponentially when idle.
The catch: zero data isolation
That shared stream is also the top critique on Hacker News — whatever one person tells the agent becomes context every other conversation can draw on. It's alpha software running real shell commands, so the docs say to sandbox it in Docker and use a spend-capped key.