autoresearch

To automate bounded mission improvement, you need an evaluator-driven loop that records per-iteration JSON results and maintains human-readable Markdown decision logs. It runs until a max-runtime or terminal condition is reached, persisting artifacts under dedicated directories for reproducibility.

Evaluator-driven single-mission improvement is an automated process that iteratively enhances one mission at a time using an evaluator to measure results. You need it when you have a defined mission specification and evaluator reference but require stateful iteration with strict stop conditions.

Yes, you must provide an existing evaluator reference and a mission specification as required inputs. The automation relies on this evaluator to assess each iteration and determine improvements before logging the decisions in Markdown and JSON formats.

You can schedule recurring automated mission iterations using Claude Code native cron functionality. This enables the evaluator-driven loop to rerun automatically while respecting the explicit max-runtime stop condition and persisting artifacts under designated directories.

The automated improvement loop is strictly limited to running one mission at a time. It enforces bounded execution with strict stop conditions like max-runtime, meaning it cannot process concurrent missions simultaneously within the same stateful loop.