Reflective prompt optimizers such as GEPA rely on analyzing an agent's failures to refine its instructions over successive iterations. But in complex agentic environments, a failure does not always point to a weakness in the prompt itself—it can instead stem from harness limitations, stochastic execution, or an unreliable automated judge. The researchers argue that such failures provide little useful learning signal, and can even be harmful, making it important to filter which tasks are fed into the optimizer in the first place.
GEPA-Curator adds two components to this type of pipeline. Task curation screens out cases likely to generate noisy or unreliable signal before they can skew optimization. Prior success retrieval selectively pulls past successful trajectories for a given task and feeds them as additional context to the optimizer, steering prompt improvement beyond simply reacting to recent failures.
The authors test the framework on four agentic benchmarks—GoBrowse, ALFWorld, AppWorld, and τ-bench—using Gemini 3.1 Pro and Claude Opus 5 as the underlying optimizer models. GEPA-Curator consistently matches or outperforms baseline methods, with reported gains of up to 20.7% over GEPA and 21.2% over ACE.
The work fits a broader push to make LLM agents self-improving without costly retraining, by acting purely on prompts and contextual information. The underlying premise—that not all agent experiences are equally valuable for reflective learning—is a reasonable one, though its robustness on real-world production tasks, beyond controlled benchmarks, remains to be seen.