Agents predict their own failure, not which fix works

Models reliably sense they're about to fail on hard problems. They can't pick the right fix, per arXiv 2608.14927, which tested 4,181 competition math problems.

That gap matters for anyone building agents that escalate to costlier reasoning modes. A simple confidence trigger hit 78% accuracy at 45,000 tokens, while a frozen router spent 71,300 tokens to reach only 73.8%.

The paper tested four protocols: solving directly, iterative self-correction, a planner-executor-reviewer setup, and multi-agent deliberation. Self-reported confidence predicted failure at 0.8847 AUROC, a strong signal. Picking which of the four protocols to escalate to was the part that broke down.

A retrospective oracle, one that could see which protocol would have worked after the fact, hit 92.4% accuracy. Neither the confidence trigger nor the frozen router came close. Between 18.5 and 28.9 points sat unclaimed by any method actually available at decision time.

Gate escalation on confidence, since models detect their own trouble well. But don't let the model choose its own collaboration structure. A model picking between self-correction, planner-executor-reviewer, and deliberation on its own buys extra tokens and nothing else.