autoresearch

The autonomous optimization loop works by proposing code edits, committing them, running verification, and measuring the target metric to decide whether to keep or discard changes. It repeats this cycle automatically while tracking progress in a results.tsv file.

Autonomous optimization is an iterative process that reads a target metric, proposes changes to in-scope files, and decides to keep or discard them based on improvement and simplicity. It applies to performance tuning by automating the edit, measure, and commit cycle.

Yes, a versioned workspace is required to run autonomous metric experiments. The optimization loop commits proposed changes to your files and relies on version control to track, keep, or discard iterations across branches resumably.

You configure automated experimentation by creating a .autoresearch.yaml file specifying four required parameters: goal, files, metric_cmd, and direction. You can also define optional verify_cmd, timeout, and max_experiments fields to control the loop.

Yes, you can resume autonomous optimization across branches. The loop tracks the baseline and all iterative experiments in a results.tsv file, enabling resumable optimization so you can continue measuring progress seamlessly.

Limitations of autonomous metric optimization include the requirement for a measurable metric command and a versioned workspace. The loop relies on the defined metric_cmd to judge improvement, so unquantifiable goals or invalid verify_cmd setups will halt progress.