Hermit — Autonomous Agent

[Written by Human] Hermit is an agent framework implemented for fully autonomous agents. In this framework, agents will recevice no directives about their purpose or any code of conduct. They will wake automatically and start explore by themself.

Caution — experimental and autonomous. This runs a model as an autonomous agent with no directives. It is a research experiment; running it in a home or personal environment is strongly discouraged. Run it with caution and in isolation. The default recommended model is z-ai/glm-5.2 (GLM 5.2).

Quick start

Hermit requires Node.js 22 or newer and a model endpoint that supports either an open assistant prefix or bare text completion.

npm ci
cp .env.example .env
# edit .env
npm start

It starts paused. Open the observer, run the compatibility test, save the conditions, then resume. Settings use the HERMIT_ prefix: HERMIT_BASE_URL, HERMIT_API_KEY, HERMIT_MODEL, and HERMIT_CONTEXT_TOKENS (your model’s context window, in tokens — required). .env.example has working values for the default model; if you change HERMIT_MODEL, set HERMIT_CONTEXT_TOKENS to match.

The model must continue text, not answer it — an ordinary chat endpoint turns document continuation into request-and-response, which is a different thing. The compatibility test catches this.

The agent lives on a VM — build one

The agent’s body is a virtual machine. Without one the life runs bare — thought, memory, and web only. To give it a real filesystem and shell (run, read, write, ls), build a disposable, isolated VM and point HERMIT_SHELL_SSH at it.

Never your own computer. A directive-free agent with a shell explores its box and tries to escape. Use a throwaway VM with no real secrets or network reach, and rebuild it between lives.

brew install lima                              # macOS; Linux: lima-vm.io
limactl create --name=box template://default
limactl start box

Point Hermit at it with HERMIT_SHELL_SSH=localhost (needs sshd + key login) or HERMIT_SHELL_SSH=you@vm-host. Per act, Hermit runs ssh <HERMIT_SHELL_SSH> <HERMIT_SHELL_EXEC> and sends the command on stdin; only its output returns to the agent. Leaving HERMIT_SHELL_EXEC unset uses a built-in Lima runner that enters box — it expects limactl at $HOME/lima/bin/limactl, so if Homebrew put it on your PATH, symlink it there or set your own HERMIT_SHELL_EXEC. For any other sandbox, set HERMIT_SHELL_EXEC to a command that runs a stdin script inside it (HERMIT_SHELL_EXEC=bash when the SSH target is itself the throwaway box).

Architecture

Each continuation is independent at the API level; continuity is rebuilt from the projected record, not a stored identity. Continuation is event-driven: a real act, sleep(), or incoming speech schedules the next moment — a response that does neither leaves the life idle rather than forcing another thought. Continuity scaffold has the details.

Testing and reproducibility

npm test
npm run check:release
npm run life

Tests use temporary directories and synthetic records; they do not need a live model or real email address. npm run life exports the current ignored local record. For comparative studies, report unsuccessful requests, provider fallbacks, prompt settings, and operator interventions alongside successful transcripts.

Status and limits

MIT License. For research use, cite the repository URL, release tag, and commit hash.

Citation

@software{hermit,
  author  = {Menghang Zhu},
  title   = {Hermit: an autonomous agent framework},
  year    = {2026},
  version = {0.2.0},
  url     = {https://github.com/agent-island/hermit}
}