Researchers Lift AI Agent Scores 21% By Ignoring Bad Runs

Researchers say GEPA-Curator made AI bots up to 21.2% better by ignoring bad test runs that taught the system the wrong lesson.

Key Takeaways

• GEPA-Curator improved agents by as much as 21.2% over one baseline and 20.7% over another
• GEPA-Curator filters tasks that produce unreliable learning signals before prompt updates
• The method was tested on GoBrowse, ALFWorld, AppWorld and τ-bench using Gemini 3.1 Pro and Claude Opus 5
• The paper argues many agent training loops treat all failures as equally informative, which can poison later updates

The researchers wrote that GEPA-Curator filters tasks that produce unreliable learning signals and retrieves successful prior trajectories before updating prompts. The paper also tested the method on GoBrowse, ALFWorld, AppWorld and τ-bench with Gemini 3.1 Pro and Claude Opus 5 as optimizer models.

Why AI Agents Learn The Wrong Thing

The core claim is simple.

Not every failed run teaches an AI agent anything useful. An agent may fail because the prompt was weak, but it may also fail because a tool broke, a judge was noisy or the environment behaved randomly.

Many agent training loops treat all failures as equally informative.

That is convenient, but it can poison later updates. If a model rewrites its prompt around a failure caused by a broken harness, it is learning from the wrong lesson.

That matters because AI agents are moving into longer tasks where errors compound.

A single bad revision can degrade performance across many future attempts.

From Reflection To Selective Memory

This paper sits inside a crowded branch of agent research. Earlier prompt optimizers such as GEPA already let an agent reflect on its own failed rollouts.

GEPA-Curator keeps that idea but adds a gatekeeper.

The gatekeeper performs two jobs. First, it filters out tasks likely to generate noisy feedback.

Second, it retrieves examples of prior success to give the optimizer better context before it changes the prompt.

For readers new to this area, a rollout is one full attempt by an AI agent to finish a task in an environment. A prompt optimizer is the system that rewrites instructions after seeing how those attempts went.

Why AI Agents Need Better Failure Triage

The practical value is not only the top-line percentage.

The paper suggests a broader rule for AI agents, that better training may come from less data if the data is cleaner. That pushes against a common instinct to feed every trace back into the loop.

Also Read: AI Cloud Firm Together Saves Costly AI Jobs By Reviving Broken Servers

It also maps onto how companies deploy agents in production.

Real software stacks contain flaky APIs, slow browsers and imperfect evaluators. If those artifacts look like model mistakes, a self-improving agent can drift away from the behavior engineers wanted.

That is one reason enterprise teams still hesitate to let agents rewrite themselves.

A system that edits its own prompt without screening the evidence can get worse while appearing busy.

What To Watch After GEPA-Curator

The next test is transfer. Research benchmarks are useful, but production teams will want to know whether curation works in customer support, coding agents and browser automation.

Those settings have more hidden failure modes than benchmark worlds do.

The paper also raises a narrower question for model labs. If curation beats naively learning from failure, future agent platforms may need built-in memory policies rather than raw log dumps.

In that sense, the result is less about one optimizer and more about a design rule.

AI agents may need selective memory, not total memory.

Why Filtering Beats More Feedback

The key mechanism in GEPA-Curator is triage. Instead of asking an optimizer to learn from every bad run, the system first asks whether the run is even trustworthy.

In AppWorld or ALFWorld, an agent can fail because it misunderstood a task, but it can also fail because the environment or judging process introduced noise. GEPA-Curator tries to separate those cases before a prompt gets rewritten.

That change matters for teams using Claude Opus 5 or Gemini 3.1 Pro as orchestration brains for AI agents. In production, logs are messy.

A browser timeout, stale tool output or inconsistent evaluator can all look like model weakness. If an agent keeps rewriting itself around those artifacts, it may wander.

What changed here versus earlier prompt optimizers such as GEPA is not raw model size.

It is the decision about which experiences count as evidence. That makes the result relevant to companies deploying AI agents now, because selective learning can be added sooner than a frontier model upgrade.

The next thing to watch is whether labs build curation into default agent frameworks rather than leaving it to researchers.

Read Next: Invesco Hands Advisers 2 Claude Tools To Build Portfolios Faster