What is a Reflection (AI Agents)?

Rather than committing to its first draft answer or first attempted solution, a reflective agent generates output, then runs a separate reasoning pass asking whether that output actually satisfies the goal, contains errors, or missed a requirement. If the self-critique surfaces a problem, the agent revises and tries again, sometimes across several rounds. This is distinct from the ReAct loop's action-observation cycle in that reflection specifically targets quality of the agent's own reasoning or output, not just the result of an external tool call. Reflection has measurably improved performance on coding and complex reasoning benchmarks, since a single pass often contains fixable errors a model can catch when explicitly asked to review its own work — much like a human proofreading their own writing catches mistakes missed while composing it. Some implementations use a separate "critic" model or prompt persona to make the self-review more rigorous than simply asking the same context to grade itself. The technique costs additional inference calls and latency, so it's typically applied selectively — for high-stakes outputs like generated code, financial calculations, or customer-facing communications — rather than on every response an agent produces.