Looped Agent Framework
_{An agent is a file.}

Looped AF lets you build contained, event-driven AI agents. Define your agent in a single config file, run it in a container and deploy it anywhere.

Contents: Demo · What's here · Docs · Contributing · License

🚧 This is alpha software 🚧

We're still in the early stages, and things are going to change and sometimes break: interfaces, config fields, defaults. If you're up for experimenting anyway, welcome. File an issue when something breaks and we'll sort it out.

Demo

Watch on youtube

The idea is that an agent is a file: one file that says the job, the model, the tools and the boundaries. Deploying it is a docker run. Agents run as long-lived services that sit in a loop; each one waits for an event (a Discord message, a webhook, a cron tick), does its one job, delivers the result and then goes idle.

handle: issue-bot     # agents name themselves; you pick the handle
description: Turns team Discord messages into GitHub issues.
model: { provider: openai-compatible, id: gpt-5.4-mini }
triggers:
  - type: discord
    channels: ["issues"]
skills:
  - ./skills/gh-issues.md
permissions:
  net: [discord.com, gateway.discord.gg, api.github.com]
  run: [gh]

What's here

• The af CLI. af init scaffolds a complete agent project with the agent file, a secrets template and a deployment shape. af validate checks it and af run runs it: without triggers you get an interactive REPL, with them the agent runs as a service.
• Triggers. Discord, Slack and Telegram (including observer agents), plus webhooks and cron.
• Capability, added deliberately. Markdown skills, MCP servers, a small native toolset and tool search, which keeps tool schemas out of a small model's context.
• Deny-by-default permissions. You allowlist the hosts, executables and paths an agent is allowed to touch. Secrets are scoped, and their values stay out of the model's context. Every permission decision lands in a SQLite audit trail.
• Docker-native deployment. We publish a hardened base image (ghcr.io/loopedautomation/agent), a one-apk add recipe for custom images and a status surface at /healthz, /runs and /audit. Deploys can be file-less, with everything passed through env vars.
• A GitHub Action. For agents that should exist for exactly one run: uses: loopedautomation/agent-framework@v0.11.1 pipes a prompt into the agent in CI and hands the reply back as a step output. See GitHub Actions.
• Budgets by default. Every run has a step cap and cheap models are the default, so you know roughly what an agent costs before you deploy it.

Docs

Published at docs.looped.sh/agent-framework and authored in docs/:

Quick start · The agent file · Triggers · Skills · Tools · Permissions · Deployment · CLI

If you'd rather start from a complete, runnable agent, the examples go from a minimal REPL bot to the Discord to GitHub gh-issues-bot (docker compose up) and agent-zero-bot, the agent that builds agents.

The runtime is Deno + TypeScript, built in the open.

Contributing

Setup, local development tasks and what CI checks are covered in CONTRIBUTING.md.

License

Apache 2.0