# AutoAgent

> An autonomous agent harness engineering tool that lets an AI meta-agent iteratively build, benchmark, and optimize agent configurations overnight without human intervention.

AutoAgent is an open-source framework for autonomous agent harness engineering — like autoresearch but for agent engineering. You give an AI meta-agent a task, and it autonomously builds and iterates on an agent harness overnight by modifying system prompts, tools, agent configuration, and orchestration. The meta-agent runs benchmarks, checks scores, keeps or discards changes, and repeats the loop until performance improves.

- **`agent.py` single-file harness** — *the entire harness under test lives in one file, containing config, tool definitions, agent registry, routing/orchestration, and the Harbor adapter boundary; the meta-agent edits this file directly.*
- **`program.md` directive** — *a Markdown file edited by the human that provides context to the meta-agent and defines the agent-engineering loop; point your coding agent at the repo and prompt it to read this file to kick off an experiment.*
- **Score-driven hill-climbing** — *every experiment produces a numeric score (0.0–1.0); the meta-agent keeps changes that improve the score and discards those that don't, following the same loop as autoresearch.*
- **Docker isolation** — *the agent runs inside a container so it cannot damage the host system, enabling safe overnight autonomous iteration.*
- **Harbor-compatible task format** — *tasks follow the Harbor benchmark format with `task.toml`, `instruction.md`, test scripts, and a Dockerfile, making it easy to port and evaluate on different datasets.*
- **Registry-driven architecture** — *agent and tool registration stay structured inside the single-file harness so it can evolve cleanly as the meta-agent iterates.*
- **Skills support** — *the agent can be equipped with Agent Skills for Context Engineering and context7 skills to improve performance on complex tasks.*
- **Quick start with `uv`** — *install dependencies with `uv sync`, set your model-provider API keys in `.env`, build the base Docker image, add tasks to `tasks/`, and run the meta-agent loop with a single prompt.*

## Features
- Autonomous agent harness engineering
- Meta-agent iterates on agent.py overnight
- Score-driven hill-climbing optimization
- Docker-isolated benchmark execution
- Harbor-compatible task format
- Single-file registry-driven harness
- program.md human-editable directive
- Agent Skills and context7 support
- Parallel task execution with configurable concurrency
- Experiment logging via results.tsv

## Integrations
Harbor benchmark framework, OpenAI API, Docker, uv (Python package manager), Agent Skills for Context Engineering, context7

## Platforms
CLI, API

## Pricing
Open Source

## Links
- Website: https://github.com/kevinrgu/autoagent
- Documentation: https://github.com/kevinrgu/autoagent/blob/main/README.md
- Repository: https://github.com/kevinrgu/autoagent
- EveryDev.ai: https://www.everydev.ai/tools/autoagent