Optimize an RLM with RLM-GEPA

RLM-GEPA improves reusable text components, usually a Skill.instructions string, using scored executions of a real PredictRLM. It is not a replacement for initial RLM design. Start only after the base RLM has a concrete signature, tools, a measurable outcome, and a working smoke path.

The system has two loops:

1. Executor loop — run each candidate RLM against train or validation examples, collecting outputs, RunTrace objects, scores, and feedback.
2. Proposer loop — use scored traces to make a surgical edit to a declared mutable component.

AgentSpec controls which transferable behaviors the proposer may improve. OptimizeConfig controls budget, concurrency, models, and search behavior. Keep those concerns separate.

Workflow