Reflection (Self-Critique)
After producing an output, the agent explicitly critiques its own work against defined criteria and revises before returning.
This page is the complete instruction page for one pattern called "Reflection (Self-Critique)." It explains the problem this pattern solves, the idea behind the solution, when you should (and should not) use it, and what happens afterward — both the good effects and the costs.
This matters because building AI agents is not just about making them clever. It is also about making them safe and predictable. Following a well-tested pattern like this one helps avoid common mistakes, and shows you exactly which safety rules and regulations it connects to, listed under "Standards Mesh" on this page.
Context
First-pass LLM outputs contain avoidable errors — unverified claims, missing steps, format violations. A critique pass catches many of them cheaply.
Problem
Single-pass generation misses errors the model could have caught with a second look.
Forces
- Quality lift vs doubled cost
- Well-defined criteria vs vague "check it"
Solution
After generation, switch the agent into critic mode with explicit, measurable criteria ("are claims grounded in retrieved sources?", "does output match the schema?"). Loop until criteria are met or a bounded iteration cap is hit.
Applicability
- High-stakes outputs
- Code generation
- Regulated-content drafting
- When quality standards are programmatically checkable
Anti-Patterns
- Vague critique prompts ("is this good?")
- Unbounded loops without exit condition
Consequences
- +Measurable quality lift on checkable outputs
- +Catches format and factuality errors
- −Doubled-or-more token cost
- −Loops risk if exit condition weak