Reflexion (Self-Critique Loops)

AssessFactorial rates Reflexion as Assess (2026). In our FlowDrop AI Bench, one architecture adds a critic that can send the draft back up to three times.

Reflexion extends an agent loop with self-review. After the agent produces a draft, a critic (the same model or another one) evaluates it and writes feedback, and the agent tries again with that feedback. The loop stops when the critic is satisfied or a retry limit is reached.

Learn more about Reflexion

Updates

Assess

In our FlowDrop AI Bench, one architecture adds a critic that can send the draft back up to three times. It was the most accurate architecture we measured (about four in five runs correct), but it cost about three and a half times as much as a fixed pipeline and took twice as long, and some of its failures were still silent. We consider it for tasks where a wrong result is expensive, and compare it with a single verifier pass first.