autoagent

Provide the GitHub URL of the agent repo and ask to run autoagent on it. The Skill clones the repo, marks the editable harness boundary, scaffolds evaluation tasks, generates a program.md directive, and hands off to the autoresearch loop for continuous keep/revert optimization.

Each task needs an instruction.md prompt and a test.sh that writes a 0.0-1.0 score to /logs/reward.txt. Scoring can be a diff against expected output or an LLM-as-judge rubric; the Skill recommends 3-10 tasks and helps generate them from example inputs and outputs.

Autoagent is intake-only: it clones the repo, marks the edit boundary, scaffolds tasks, and writes program.md. Autoresearch runs the actual optimization loop with keep/revert decisions, failure taxonomy, and trace logging via its agent-harness domain.

Use the original AutoAgent directly when you need Harbor container isolation, ATIF trajectory output, or the OpenAI Agents SDK specifically. Use this wrapper when you want the autoresearch loop's trace management, stuck detection, and domain routing.

Benchmark replay only supports a fast single-session keep/revert loop, which risks overfitting to the specific tasks. The Skill includes an overfitting guard, but optimizing for real-world success rates requires autoresearch's epoch-based loop instead.