Creativity · Agent Protocol

Reflection Agent Pattern

Reflection is a self-critique pattern: after producing a first answer, the agent is prompted to review its own work, identify errors or omissions, and revise. A common variant uses two agents — a generator and a reviewer — with the reviewer playing the critical role. Papers like Reflexion and Self-Refine showed that this simple loop can noticeably improve quality on code, writing, and reasoning tasks.

Protocol facts

Sponsor
open research pattern
Status
stable
Spec
https://arxiv.org/abs/2303.11366
Interop with
LangGraph, AutoGen, CrewAI, A2A

Frequently asked questions

Does reflection always help?

Usually, but not always. Reflection helps most on tasks with verifiable structure — code, math, factual writing — where a reviewer can spot concrete errors. On open-ended creative tasks the second pass sometimes makes things blander.

Is reflection the same as chain-of-thought?

No. Chain-of-thought is thinking in the middle of one answer. Reflection is producing an answer, then explicitly stepping back to critique it — usually with a separate prompt or agent.

One agent or two?

Both work. Two-agent reflection (generator + reviewer) tends to catch more errors because the reviewer isn't anchored by having written the draft, but it costs more.

Sources

  1. Shinn et al. — Reflexion — accessed 2026-04-20