GEPA

GEPA is an open-source research framework for improving AI systems through reflective optimization. Developers supply something to optimize and an evaluator that measures its performance. Its optimize_anything API extends this approach beyond prompts to code, agent architectures, configurations and vector graphics. It supports solving individual problems, sharing improvements across related tasks, and learning artifacts that generalize to unseen examples.

Developed at the UC Berkeley Sky Computing Lab through a collaboration between UC Berkeley, MIT, Stanford and Databricks, GEPA emerged from prompt-optimization research published in 2025. Its Genetic-Pareto approach uses an LLM to inspect execution traces, diagnose failures and propose revisions. Rather than retaining only the candidate with the highest average score, it preserves candidates that excel on different examples and combines complementary improvements. This makes diagnostic feedback—such as errors, tool outputs and reasoning traces—part of the optimization process.

The implementation integrates with DSPy, MLflow, Pydantic AI and Comet ML’s Opik, letting developers optimize within existing AI workflows. Prompt optimization can work through model APIs without access to model weights. In 2026, optimize_anything added interchangeable optimization engines and composable pipelines: developers can run several engines, retain the strongest result, or pass a candidate between engines for further improvement.