# GEPA-Curator: Curated Learning for Self Improving Agents

Source: [arXiv](https://arxiv.org/abs/2610.37651v1)  
Feed7 permalink: https://feed7.dev/p/2610-37651v1-1usz0jw  
Published: 2026-10-27T17:58:11.000Z  
Trust: Needs Review (needs_review)

## Why Included

GEPA-Curator filters noisy agent failures and retrieves relevant successful runs before prompt optimization, improving results over two baselines across four environments.

## Source Summary

GEPA-Curator adds **task curation** to filter experiences with unreliable learning signals and **prior success retrieval** to add relevant successful trajectories as context. It was evaluated on four environments with Gemini 3.1 Pro and Claude Opus 5 as optimizer models.

## Practical Implication

If you optimize prompts from coding-agent runs, separate useful failures from harness issues, stochastic outcomes, and noisy judges; consider reusing successful trajectories for similar tasks.

## Agent-Ready Context

GEPA-Curator adds **task curation** to filter experiences with unreliable learning signals and **prior success retrieval** to add relevant successful trajectories as context. It was evaluated on four environments with Gemini 3.1 Pro and Claude Opus 5 as optimizer models.

If you optimize prompts from coding-agent runs, separate useful failures from harness issues, stochastic outcomes, and noisy judges; consider reusing successful trajectories for similar tasks.

The paper reports gains of up to **20.7% over GEPA** and **21.2% over ACE**, but the material does not specify how these gains vary by environment or how much curation costs.

## Connected Context

Feed7 judgment across 944 accumulated Signals:

GEPA-Curator adds two controls to prompt optimization: filter experiences with unreliable learning signals and retrieve relevant successful trajectories. The reported gains are bounded by four environments and two optimizer models; the supplied record leaves variation by environment and curation cost unspecified.

- [Improving Agents is a Data Mining Problem — Vivek Trivedy, LangChain](https://feed7.dev/p/improving-agents-is-a-data-mining-problem-vivek-trivedy-langchain-1cgibr8) — LangChain describes mining production traces for failures and replayable evals; GEPA-Curator adds a way to screen those experiences for usable learning signals and retrieve relevant successes.
- [The Rise of Verbal Reinforcement Learning](https://feed7.dev/p/2609-01597v1-1yaem2a) — The survey distinguishes feedback retained as context from other learning interventions; GEPA-Curator's successful-trajectory retrieval is a context intervention, while its task curation filters optimizer inputs.
- [FOCUS: Active Critic Learning for Long-Horizon Language Model Agents](https://feed7.dev/p/2610-35801v1-1bbefst) — FOCUS selectively samples states to improve long-horizon critic estimates, while GEPA-Curator filters whole experiences for reliable learning signals; both address where optimization effort should be spent, at different stages.

## Context Map

- Layer: agent
- Domains: research
- Topics: harness-engineering, prompting, agent-evals

## Uncertainty

- The paper reports gains of up to **20.7% over GEPA** and **21.2% over ACE**, but the material does not specify how these gains vary by environment or how much curation costs.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.