Self Improve

This is an autonomous code optimization loop that runs tournament selection on competing improvement strategies. It spawns specialized agents (researcher, planner, architect, critic, executor) to analyze your codebase, generate hypotheses, implement changes, and benchmark results without stopping to ask permission between iterations. The system runs inside a target repo you specify, executes your benchmark command repeatedly, and keeps the best performing code through a git-based tournament. It includes guardrails against modifying benchmark code and circuit breakers for regressions. You'd use this when you have a measurable performance target and want to throw compute at the problem overnight. The trust model is explicit: you confirm the repo path and benchmark command once, then it runs hot until plateau or target.