self-improve

Autonomous code improvement loops are orchestrated by coordinating research, planning, execution, and evaluation. The process uses tournament-style selection among multiple agents to iteratively optimize code while enforcing safety guardrails and sealed-file constraints.

Tournament-style selection runs multiple autonomous agents that generate competing code improvements. The system evaluates these iterations against benchmarks, selecting the best-performing changes to merge while discarding less effective experiments.

Yes, git worktree management creates isolated environments for autonomous experiments. This ensures iterative code changes are tested safely without affecting the main branch and maintains a historical record of all optimization trials.

You need jq and git installed. You must also initialize the target repository path, define a clear optimization goal via the goal-clarifier, and generate a benchmark to evaluate improvements before starting the iterative process.

Safety guardrails enforce sealed-file constraints and safelists to prevent unauthorized modifications. Trust gates evaluate autonomous changes against stop conditions before allowing merges, ensuring all code experiments remain auditable and reversible.

Yes, the system features robust state tracking and resumability. If an autonomous code improvement iteration is interrupted, you can resume progress from the last saved state without losing experimental data or historical records.