Creativity · Agent Protocol

Agent Self-Critique Pattern

In the self-critique pattern, the agent does not return its first draft. Instead it critiques that draft against a rubric — checking for hallucinations, missed requirements, style issues, or security bugs — and either revises in place or marks the draft for rework. Adds latency and cost, but often halves the error rate on complex tasks like code generation and legal drafting.

Protocol facts

Sponsor
Research community
Status
stable
Interop with
LangGraph, AutoGen, CrewAI, any agent framework

Frequently asked questions

When should I use self-critique?

When the cost of a wrong answer is high (production code, legal drafts, customer-facing comms) and you can afford 2–3x the latency and token cost. Skip it for low-stakes, high-volume tasks.

Does the same model critique its own work?

It can, but you get better results from a distinct critic — a stronger model, a different system prompt, or a specialized verifier. Self-critique by the same model catches fewer errors than cross-model critique.

What's the downside?

Latency, cost, and occasional over-correction where the critique introduces new errors. Cap the refinement loop at 2–3 iterations and compare final outputs against a held-out eval set.

Sources

  1. Reflexion paper — accessed 2026-04-20
  2. Self-Refine paper — accessed 2026-04-20