autoresearch-fleet

You can automate code optimization by running an autonomous loop that iteratively edits a file, evaluates it with a fixed harness, and commits metric improvements while reverting regressions. This requires a git-initialized workdir and a configuration file defining the evaluation command.

Yes, autonomous research loops support provider-flexible orchestration with claude, codex, or pi providers. The orchestrator manages stop conditions under per-iteration budget caps, allowing long-running experiments across different model providers.

Plateau-triggered web search is a mechanism that detects consecutive discarded or crashed iterations and switches to a web-search-augmented prompt. This breaks through knowledge ceilings when the autonomous loop stalls during code optimization.

To set up autonomous benchmarking, prepare a fleet root directory containing a configuration file with your eval command, metric extraction, and search settings, alongside a program instruction file. You need a git-initialized workdir to maintain the experimental trail.

Use git as a state machine for algorithm refactoring when you need a clean experimental trail during continual exploration. Improvements advance history while regressions trigger a hard reset, ensuring only validated metric improvements are kept in the codebase.

Autonomous optimization loops are limited to scenarios with deterministic benchmarks and fast testable metrics, such as latency or loss tuning. They require defined budget caps and stop conditions, and depend on a fixed evaluation harness to validate metric improvements.