self-improve

Automated code improvement uses a YAML-configured loop to iteratively generate hypotheses, execute changes in isolated git worktrees, and select winners via benchmark scoring. It eliminates manual trial-and-error by validating performance gains against a deterministic evaluation command.

Tournament selection for autonomous codebase optimization is a process where multiple candidate code changes are executed and scored against a benchmark. The highest-scoring changes are selected as winners, driving systematic evolutionary improvement until explicit stop conditions are met.

To safely run autonomous code modification, execute changes in isolated git worktrees and enforce sealed-file constraints using a validation script. This strategy isolates experiments and prevents unintended edits to the repository while testing performance optimization hypotheses.

Yes, you can use this autonomous evolution loop for repository-level performance regression testing by defining a benchmark command and stop conditions. It repeatedly evaluates measurable performance metrics in a trusted local environment to search for verifiable gains.

Yes, you need a YAML configuration to start the loop, as it reads goal, benchmark, and harness settings. This configuration drives the orchestrator, delegates tasks to agents, and enforces sealed-file and plan/result JSON schemas throughout the improvement lifecycle.

A limitation of using benchmark scoring for code optimization is that it requires a deterministic evaluation command and measurable metrics. Without explicit termination conditions and a trusted local environment, the loop cannot safely validate performance gains or stop correctly.