AutoAgent

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.