Frontier models optimize endlessly but never question the problem

We deployed frontier models as R&D agents last month. The results surprised us — not because they failed, but because of how they failed.

Everyone talks about agents being creative or autonomous. The reality is more mundane: they're excellent optimizers that hit walls you don't expect.

The optimization trap

Our agents could iterate on solutions endlessly. Give them a machine learning task, and they'd run 47 variations of gradient descent with different learning rates. Ask them to improve API performance, and they'd try every caching strategy in the textbook.

But ask them to question the problem itself? Crickets.

Out of 252 solutions across our evaluation, exactly 3 qualified as genuinely novel approaches. The rest were sophisticated combinations of existing techniques — which is useful, but not the breakthrough thinking we expected.

The reliability cliff

Here's what caught us off-guard: two agents could achieve identical final results but fail completely differently. Agent A would frame problems incorrectly from the start. Agent B would execute flawlessly but misinterpret feedback loops.

We started tracking failure modes instead of just success rates. The pattern became clear: consistent execution separates production agents from demo agents, not peak performance.

// Failure mode tracking we added
{
  "solution_framing": "incorrect_assumptions",
  "execution_quality": "high",
  "feedback_interpretation": "missed_key_signals",
  "final_score": 0.73,
  "reliability_score": 0.31
}

The experience poisoning problem

Accumulated experience helped agents preserve useful discoveries between sessions. But it also carried forward misleading conclusions that anchored them to local optima.

Our agents would "remember" that approach X failed in January and avoid it in March — even when the constraints had completely changed. Memory without context decay becomes a liability.

Warning: Experience reuse is a double-edged sword. Agents that learn too well become prisoners of their own history.

The harness matters more than the model

The biggest surprise: automated harness optimization produced gains that transferred across tasks and even between different models. The infrastructure around the agent — how it receives feedback, manages iterations, handles failures — matters more than raw intelligence.

We're not building better researchers. We're building better research environments.

What this means for your agents

Stop optimizing for creativity. Start optimizing for reliability. Your agent doesn't need to invent new paradigms — it needs to consistently execute the paradigms that already work.

Track failure modes, not just success rates. Build harnesses that can transfer between agents. And remember: experience without decay becomes baggage.