agents · review · first seen Oct 28, 09:33 UTC

GEPA-Curator filters agent experiences for prompt optimization

GEPA-Curator was announced as a way to select which agent experiences inform prompt optimization and to supply past successes as context. Builders should care because a failed rollout may reflect a noisy judge or harness limitation rather than a prompt weakness.

L0 · primary1 source4 claims checked4 min read

What shipped

Despite this section’s name, the supplied source establishes an announcement, not a shipped tool. GEPA-Curator is a framework for choosing which agent experiences shape prompt optimization. Its premise is that an optimizer learning from failed rollouts needs to distinguish useful feedback from failures that may not say much about the prompt. The excerpt names prompt weaknesses, harness limitations, stochastic execution, and noisy judges as possible causes of failure; it does not say that the curator can diagnose each cause with certainty. [1][2]

The announced framework has two components. Task curation filters tasks likely to provide noisy or unreliable learning signals. Prior success retrieval selects past successful trajectories for a task and gives them to the prompt optimizer as additional context. These are distinct decisions: whether a task should contribute a learning signal, and which earlier successes might help the optimizer interpret a task. The source does not specify a filtering threshold or a retrieval interface. [1][2]

The authors report evaluating GEPA-Curator on GoBrowse, ALFWorld, AppWorld, and τ-bench, with Gemini 3.1 Pro and Claude Opus 5 as optimizer models. That is evidence of an evaluation across the named workloads, not evidence in the supplied text that the framework is available to install. The excerpt cuts off during a performance statement, so it does not provide a complete result to report. [1]

Why it matters if you build

If you already use failed agent rollouts to revise prompts, the useful change is to inspect the input to that revision, not just the proposed revision. A failure caused by a harness limit or an unreliable judgment may prompt an edit that addresses the wrong problem. GEPA-Curator’s announced task curation addresses that risk by filtering tasks likely to carry weak learning signals. The claim is about selecting experiences; it does not establish that every retained failure identifies a prompt weakness. [1][2]

Prior success retrieval offers a second workflow choice: show the optimizer relevant successful trajectories alongside the task being considered. For a builder, this suggests keeping successful runs available for review rather than treating only failures as optimization material. The source says the retrieval is selective, but does not explain its selection rule or what context length it requires. Any storage, review, or optimizer-context cost in your own workflow therefore remains something to measure, not a published figure in these claims. [1]

The named evaluations give builders a place to compare their workload with the authors’ test settings. They do not, on their own, establish a benefit on an unlisted environment or with a different optimizer model. Nor does the supplied excerpt provide a complete numerical comparison, so a decision to replace a working optimization loop would need more evidence than these claims contain. [1]

How to try it today

1. List the failed agent tasks you currently feed into prompt revision. Mark any whose outcome could reflect a harness limit, stochastic run, or noisy judgment.
2. Set those uncertain tasks aside for separate review instead of treating every failure as a prompt-edit instruction. This is a manual test of the curation idea, not a stated GEPA-Curator feature.
3. For one task you do review, find a past successful trajectory and present it as additional context when considering a prompt change.
4. Compare the prompt decision with and without that success context; record whether it changes your interpretation. Do not treat a changed interpretation as proof of improved performance.
5. Check whether your agent workload resembles any of GoBrowse, ALFWorld, AppWorld, or τ-bench before using the reported evaluation as a guide.

What we checked

Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.

1. announced L0 · primary

   GEPA-Curator selects which agent experiences shape prompt optimization.

   per arXiv cs.CL/cs.AI/cs.LG (L0)
2. announced L0 · primary

   GEPA-Curator filters tasks likely to provide noisy or unreliable learning signals.

   per arXiv cs.CL/cs.AI/cs.LG (L0)
3. announced L0 · primary

   GEPA-Curator retrieves past successful trajectories as context for the prompt optimizer.

   per arXiv cs.CL/cs.AI/cs.LG (L0)
4. announced L0 · primary

   GEPA-Curator was evaluated on GoBrowse, ALFWorld, AppWorld, and τ-bench.

   per arXiv cs.CL/cs.AI/cs.LG (L0)

What is not confirmed

The supplied material does not establish that GEPA-Curator shipped as a usable package, service, or integration. It gives no installation steps, task-filtering criteria, retrieval settings, or operating cost. Treat the two components as an announced framework, not a ready-to-adopt interface. [1][2][3]

The excerpt names evaluation settings but ends before completing its performance statement. It therefore does not establish a numerical gain, a consistent gain across all four workloads, or how either component performs alone. Do not use it to promise an improvement for your agent. [1][2][3]

Who should act, who should wait

Act now if you already maintain an agent prompt-optimization loop and can review its failed tasks and past successful trajectories. Try the selection and context ideas within that existing workflow, while keeping the result separate from any claim that GEPA-Curator itself is deployable. The immediate decision is whether your optimizer is learning from experiences you trust. [1][2][3]

Wait before committing to GEPA-Curator as a dependency or claiming a measured benefit. The supplied material announces the framework and names evaluation workloads, but does not provide deployment instructions or a complete performance result. Builders who need either should seek those details before changing a production process. [1][2]

Sources

• GEPA-Curator: Curated Learning for Self Improving Agents — arXiv cs.CL/cs.AI/cs.LG L0 · primary