Core agents
What is Reflection?
Reflection is a later pass that inspects an artifact and proposes a correction.
It is a workflow stage with inputs and outputs, not a window into private reasoning.
Why it matters
Reflection is a second pass where the agent inspects its own output and proposes corrections. It catches errors that the first pass missed — factual mistakes, incomplete answers, policy violations.
Reflection is a workflow stage, not a personality trait. It has inputs (the artifact to review) and outputs (the critique or corrected artifact).
How to implement reflection
The simplest reflection pattern is: generate an answer, then prompt the same or different model to review that answer against criteria. The reviewer returns either "pass" or a list of issues. If issues are found, the agent re-generates.
Production systems limit reflection to one or two passes to prevent infinite self-criticism loops.
Key takeaways
- 1Reflection is a designed workflow stage with clear inputs and outputs.
- 2Limit reflection passes to prevent infinite self-criticism.
- 3Use reflection for high-stakes outputs, not every response.
Common mistakes
- ✕Treating reflection as free — each pass costs tokens and latency.
- ✕Reflecting without criteria, which produces vague self-congratulation.
Related terms
Concept neighborhood
Terms linked from Reflection in the glossary graph.
- Reflection
- Planning
- Evaluator Agent