Optimize

An autonomous optimization loop iteratively improves a target by cycling through seven phases: OBSERVE, THINK, PLAN, BUILD, EXECUTE, VERIFY, and LEARN. You can run it using a measurable metric or LLM-guided evaluation to systematically search and apply improvements.

Yes, LLM-guided evaluation supports optimizing qualitative targets like skills, prompts, and agent workflows. Eval mode applies automated evaluation criteria to iteratively refine and enhance the quality of your autonomous agent workflows.

Autonomous code optimization runs sandboxed experiments by applying CLI arguments like --target and --metric. This isolates tests during the EXECUTE and VERIFY phases, ensuring safe iterative improvements without destabilizing your main environment.

Metric mode targets codebases using quantifiable metrics like performance or bundle size, while eval mode targets qualitative evaluations of skills and prompts. Both modes drive the seven-phase autonomous optimization loop to improve the target.

An interrupted autonomous optimization loop can be resumed using built-in status checks and resume functionality. This allows the workflow to continue from its last known state without losing prior experimental progress or learned data.

You can define budget constraints and limits using CLI arguments like --budget and --max-experiments. These parameters restrict the number of iterative testing cycles the autonomous optimization loop executes before stopping.