AI agents get better at fixing their own mistakes — when they can explain why they failed

Researchers built a system that lets AI agents categorize their failures into ten specific types instead of just knowing they failed, then use that diagnosis to pick better strategies next time. In practice, this means an AI assistant doing multi-step tasks can now learn from what went wrong — not just whether it went wrong — and apply that learning to similar problems without starting from scratch.

For years, AI agents were stuck with a crude feedback loop: task succeeds or fails, try something else. This system actually names what broke — parsing errors, strategy errors, execution errors — and stores causal relationships (if this condition, then that action) so the agent can retrieve relevant past solutions. The structural change is simple but real: an agent can now inherit experience instead of repeating the same mistakes.