Can AI agents correct their own mistakes?

Can an AI agent reliably catch its own mistakes?

No. The same model that ignores an error in its own reasoning trace will flag that error when it appears as an external claim. In the 2026 study the correction rate rose 23 to 93 points just by relabeling the identical claim from the agent's own thought to an external role. The ability is there; the prompt structure suppresses it.

Why doesn't a self-reflection or double-check step fix it?

Because a self-reflection step routes the claim back through the agent's own role, the exact condition where correction fails. You get a paragraph that says it looks correct without a real catch. The model trusts its own prior turns, so to trigger a correction the claim has to arrive as something it did not say.

Will a bigger model fix self-correction?

Unlikely on its own. The effect held across seven model families and 13 model-domain cells, with 10 of 13 significant at p < 0.001. It is a structural feature of how the chat template tags roles, so a separate reviewer beats a bigger single model that still grades its own work.

What actually makes agents catch their errors in production?

Put the claim in front of a different role: a separate reviewer agent, a tool or eval response, or a canonical memory layer the agent reads as external input. Separate the doer from the checker and gate on an external eval, not a self-vote. That is an operating choice, not a prompt tweak.