headlong

██  ██ █████  ████  █████  ██     ████  ██  ██  ████               ██
██  ██ ██    ██  ██ ██  ██ ██    ██  ██ ███ ██ ██           ██   ██
██▀▀██ ████  ██████ ██  ██ ██    ██  ██ ██▀███ ██ ███     ██   ██
██  ██ ██    ██  ██ ██  ██ ██    ██  ██ ██ ▀██ ██  ██   ██   ██
██  ██ █████ ██  ██ █████  █████  ████  ██  ██  █████      ██

Headlong is an open source agent microharness, a complete agent harness
with a core of less than 10K lines of Bash. Headlong's defining feature is
persistent agency. Your agent keeps thinking between external
interactions in a self-guided loop inspired by human inner monologue. A
message from a human doesn't start a session. It lands in the agent's
thought stream as one more observation, and the agent decides if and when
to respond. You give your agent a name and a personality, and it sets its
own interests and priorities, starts its own projects, and pings you when
it has something to say.

A Headlong agent is also built to be shared. A whole team can talk to one
agent over Slack, Telegram, and a chat app, and every conversation lands
in the agent's single stream of thoughts. The agent follows what different
people are working on, connects them, and pings whoever seems most
relevant. Sharing one agent is fun, because it behaves more like a person
than a service.

At the heart of Headlong is shellm, a Bash implementation of a
recursive language model (RLM).
The agent thinks by writing shell commands, running them, and reading the
output. No tool system besides Bash is needed.

Get started

One line installs everything, interviews you to bring a Headlong agent to
life, and opens a dashboard where you can watch its mind run:

curl -fsSL https://headlong.ai/install.sh | bash

You'll need bash 3.2+, git, curl, jq, and an LLM API key (Anthropic,
OpenAI, Gemini, or OpenRouter); the dashboard also needs
uv and bun or node, and the installer offers
to fetch those.

Headlong is alpha research software. Use a dedicated, spend-capped key, because
your agent runs real shell commands and thinks around the clock. With Docker
running, the installer offers to keep the whole agent in a container, or to
install on your machine with the agent's commands sandboxed in a container
(an unsandboxed host install exists too, behind an explicit yes, and is not
recommended).
Without Docker the commands would run directly on your machine as you, so the
installer stops and asks for an explicit yes before setting that up. How much
the background thinking costs depends on how quickly
the agent loops and which model backs it. The rate of thinking backs off
exponentially when nobody is talking to the agent and resets the moment a
message arrives. At the settings we run our agent with, it comes to $1 to $2
an hour.

The agent's name becomes a command:

ada hello            # one message, wait for the reply
ada                  # chat
ada stop / ada start # pause / resume its mind
ada dash             # open the dashboard
ada bugreport        # bundle logs + trajectory (keys scrubbed) for a bug report

headlong-killall stops every Headlong process on the machine if you need a
panic button. curl -fsSL https://headlong.ai/status.sh | bash shows what is
installed and running; curl -fsSL https://headlong.ai/uninstall.sh | bash
removes it all (details in
docs/install.md).

The container flow the installer offers is this, and you can also run it
yourself:

docker run -it --name headlong --restart unless-stopped -p 8080:8080 buildpack-deps:curl \
  bash -c 'curl -fsSL https://headlong.ai/install.sh | bash; exec bash'

Details, non-interactive/CI installs, and installing from a checkout are
in docs/install.md.

Key ideas

• Persistent agency. Most harnesses are reactive, or wake on a
  schedule to run a fixed checklist. A Headlong agent is never asleep and
  there is no checklist. It keeps generating thoughts about whatever it
  decides is interesting, even when there is no external input. Messages
  from Slack, Telegram, or the chat app are injected into the thought
  stream as observations, and the agent decides if and when to respond.
  Classic turn-taking request/response mode works too.
• Multi-player fun. One agent, one mind, many people. There are no
  per-user sessions; the agent experiences all of its conversations in a
  single timeline and decides who to reply to and when. That single
  stream also means no hard walls between people: assume anything you
  tell the agent is shared with everyone who talks to it.
• Built around Ken Thompson's philosophy. The core tooling is a
  handful of small Bash executables (shellm, traj, llm, context,
  mem, skills, ...), each doing one thing well and composing through
  pipes, files, and environment variables. The model writes shell
  commands, so curl is the HTTP client and jq is the JSON processor.
• An agent's trajectory is a DAG of jsonl files with fork and merge.
  An agent has access to everything it has thought and done, and the
  tooling to explore it down to any single step.
• Context is a projection of the trajectory. Nothing is compacted
  away in place. Compaction and agent introspection operate on the same
  files with the same tools.
• Tiered context compaction. The entire trajectory stays in context
  at exponentially decaying resolution. Recent entries appear verbatim,
  and older entries are progressively summarized. The tiers act as an
  index, so the agent can retrieve raw entries when it needs them.
• Subagents see their ancestors' trajectories. A subagent can see why
  it was created, what the parent already tried, and how it fits into the
  big picture.
• Docker by default. Generated code sandboxes itself into a container
  whenever Docker is available, and container reuse keeps restarts cheap.
  Local mode works too.
• Self-improvement by fork, test, merge. An agent forks the Headlong
  codebase (and optionally its own trajectory), changes something, and
  runs. Merge the change back if it worked, or discard the agent and its
  changes if it didn't. No rollback machinery is needed. The agent we run
  at Laude works in its own fork of this repo, and we have pulled over 50
  of its commits back into main.

The full backstory and design philosophy are in
philosophy.md.

The tools

To make a minimal agent, you need:

• a loop that repeatedly generates the next thought (thinkers, which
  calls llm),
• a way for a thought to reason and act (shellm, with Bash as the only
  tool),
• a way to record the agent's trajectory, its life so far (traj), and
• a way to turn that trajectory into the context for the next call into
  the LLM (context).

Headlong also gives an agent a few convenience tools, such as a way to
distill and codify its experience (mem) and a way to save and reuse
procedures for specialized tasks (skills). The core is the tools the
running mind executes, the executables in bin/ plus the thought
processes in thinkers/, and it comes to 9.8K lines by cloc's count. A
harness this small can be read end to end, and it is easy to modify and
experiment with.

Everything you run around the mind lives in tools/:

Learn more

• philosophy.md — the case for applying Ken Thompson's
  philosophy to agent microharnesses, and the full design story
• docs/shellm.md — the shellm engine reference: the
  loop, context passing, Docker sandboxing, envs, the llm tool, options
• docs/install.md — every install variant, including
  CI/non-interactive and long-lived Docker
• AGENTS.md — operating a running identity (for humans and
  coding agents): paths, logs, health checks, sharp edges
• web/, slack/,
  telegram/ — the dashboard and the chat bridges
• deploy/ — running an agent on a dedicated box
  (systemd units, terraform, operations)

Acknowledgements

The recursive language model idea in shellm comes in part from the
Recursive LLM experiment (April
2023) and from Alex Zhang's Recursive LM
(RLM) project (October
2025). The continuous thinking behind Headlong's persistent agency — and
its name — come from the Headlong
research project.

License

Apache 2.0. Copyright 2026 Laude Institute.