harness:evolve

Agent optimization is automated by orchestrating propose-evaluate-iterate loops that use LangSmith for evaluation and Git worktrees for isolation. This structured iteration enables teams to systematically evolve agent performance across experiments.

Evolving agent performance is structured through configurable modes like light, balanced, or heavy iterations. These modes coordinate multi-stage propose, evaluate, and merge cycles with baseline scoring and traceability for artifact management.

Yes, automated agent evaluation works with Git worktrees to provide isolation during iteration loops. This ensures that each proposed change is evaluated independently without interfering with the main codebase or parallel experiments.

To start evolving agents, run the setup harness to generate the required .evolver.json configuration file. This enforces the necessary environment setup and provides parameters for iterations, interactive prompts, and preflight checks.

Yes, you can run preflight checks and baseline scoring before starting automated agent iterations. The workflow validates the environment and establishes baseline metrics to measure subsequent performance improvements accurately.

Automated agent iteration loops require a .evolver.json configuration file and LangSmith evaluation backend integration. Teams must manage Git worktree artifacts and coordinate multi-stage cycles, which may increase infrastructure complexity at scale.