GEPA Optimization

GEPA (Generative Evolutionary Prompt Approximation) is Kelet’s validated prompt improvement engine. Where a prompt patch is a single LLM-generated suggestion, GEPA runs an iterative search — testing variations, measuring quality against your real failure patterns, and converging on a fix with evidence it works.

How it works

1. Kelet takes the root cause identified for an issue and the failing session traces
2. Generates candidate prompt variations
3. Evaluates each variation against your synthetic evaluators and historical signals
4. Selects the best performer, generates new variations from it, and repeats
5. Returns the winning prompt with before/after quality metrics

The result isn’t a suggestion — it’s a validated improvement with a measurable quality delta.

Running GEPA

1. Open Findings in the console and select a Finding
2. Click the chevron on the Prompt Patch split button → select GEPA Optimization
3. Review the intro — GEPA explains what it will fix and how it measures success
4. Click Start Optimization
5. Watch the live progress — each generation shows the current best score
6. Review the results: before/after metrics, the optimization journey, and the winning prompt
7. Copy the prompt and apply it to your agent

What GEPA needs

GEPA uses your synthetic evaluators to measure quality during the search. The more relevant your evaluators are to the failure mode, the better the result. If you haven’t set up synthetic evaluators yet, do that first — see Synthetic Signals.

GEPA vs. prompt patch

	Prompt patch	GEPA
Speed	Instant	Minutes
Method	Single LLM suggestion	Evolutionary search
Validation	None	Tested against your evaluators
Best for	Quick fix to explore	High-confidence fix to deploy