autoagent

Why we included this project

Anyone who has spent an evening hand-tuning an agent harness, one benchmark run at a time, will recognize the loop AutoAgent automates. The meta-agent edits the harness's single-file agent.py, runs the Harbor-based task suite, reads the numeric score, and keeps or discards each change on its own. You steer the whole process through program.md instead of touching Python directly, which keeps experiments reproducible and easy to hand off. For engineers who already build agents, it is a practical way to get an automated hill-climbing loop over their own evaluation tasks, and the Docker isolation plus score-driven keep/discard logic makes it safe to run unattended. The repo ships without tasks, so expect to bring your own Harbor-format benchmarks before the loop has anything to optimize.