Reflexion

Reflexion is an agent pattern introduced at NeurIPS 2023. Instead of running a single reasoning loop and stopping, the agent writes a natural-language critique of its own previous attempt, stores it in memory, and uses it as context for the next try. The motivation was a known weakness of ReAct-style loops: making the same mistake repeatedly.

The flow is straightforward — act, evaluate the result, write a self-reflection note, then attempt again with that note in context. Reported gains over plain ReAct on coding and math-reasoning benchmarks are what made the pattern stick.

The practical caveat is that the self-critique is not always right. A bad reflection can make the next attempt worse, so production systems usually pair Reflexion with an external check — a test runner, a judge model, a verifier — to confirm whether the reflection actually improved things.