Imagine you're coaching a basketball team by reviewing game tape. After every loss, you sit the team down and dissect what went wrong. But some losses weren't the team's fault — a bad referee call, a slippery court, a scorekeeper's error. If you coach your players to 'fix' those problems, you'll make them worse, not better. They'll start second-guessing moves that were actually correct. GEPA-Curator is the assistant coach who watches the tape first and says: skip that one, the floor was wet — here are the losses we can actually learn from. The committed claim: not all agentic failures carry useful learning signal, and blindly feeding failed rollouts into a prompt optimizer actively degrades performance. GEPA-Curator adds a curation layer — task filtering plus retrieval of prior successes — on top of the existing GEPA reflective prompt optimization loop. The gains are real: up to 20.7% over vanilla GEPA and 21.2% over ACE across four diverse benchmarks (GoBrowse, ALFWorld, AppWorld, τ-bench). This paper sits squarely in the growing subfield of self-improving LLM agents — systems where the model iterates on its own prompts or policies through reflective loops. The direct predecessors are GEPA (reflective prompt optimization from failed rollouts) and ACE (a similar self-improvement baseline). The core insight isn't architecturally novel — it's a data-quality argument applied to prompt optimization. If you've worked in ML at all, 'garbage in, garbage out' is not a surprise. What's useful here is the decomposition of failure sources: prompt weaknesses (learnable), harness limitations (not learnable), stochastic execution noise (not learnable), and noisy judges (actively misleading). The authors argue — correctly — that conflating these failure modes poisons the optimization signal. The architecture is straightforward: GEPA-Curator wraps around any reflective prompt optimizer with two modules. Task curation filters incoming failed rollouts based on estimated noise/reliability signals before they reach the optimizer. Prior success retrieval pulls successful trajectories for the same task from a memory bank and injects them as additional context, giving the optimizer a positive reference frame alongside the failure. This is not a new training paradigm — it's a preprocessing and retrieval augmentation layer. The compute overhead is minimal; the intellectual contribution is recognizing that curation matters more than optimization sophistication in noisy agentic settings. Integrity is mixed. The evaluation spans four benchmarks (GoBrowse, ALFWorld, AppWorld, τ-bench), which is broader than many agent papers. Both Gemini 3.1 Pro and Claude Opus 5 serve as optimizer models, reducing the single-model-dependency risk. The paper reports that GEPA-Curator 'matches or exceeds the mean performance of both baselines' — note the careful language. 'Matches or exceeds the mean' is not the same as 'always wins.' The up-to figures (20.7%, 21.2%) are ceiling gains on specific benchmarks, not average improvements. No code release or pre-registration is mentioned in the abstract. The baselines (GEPA, ACE) are current and relevant, not stale. The milestone question is where this gets interesting for practitioners. The real unlock isn't a specific number — it's the principle that curation quality scales the ceiling of self-improvement loops. As agentic workloads move toward longer multi-turn interactions (dozens or hundreds of steps), the noise-to-signal ratio in failure rollouts will only get worse. The practical threshold to watch: can curated self-improvement close the gap to human-level prompt engineering on complex agentic tasks? We're nowhere near that yet, but this paper argues the bottleneck is data quality, not optimizer quality. The obvious missing experiment is ablation across failure types. The paper taxonomizes failures (prompt weakness, harness limitation, stochastic noise, noisy judges) but the abstract doesn't report per-category filtering impact. How much of the gain comes from filtering noisy-judge failures versus stochastic execution noise? This decomposition would tell practitioners where to focus their own curation efforts. My honest read: this is either in the full paper and just not in the abstract, or it's being saved for the next paper. The framework practically invites this ablation.
