Reflection

Process agents published

Also known as: Self-Reflection, Self-Critique, Meta-Cognition

Definition

The process by which an agent evaluates its own outputs, reasoning, or behavior and uses that evaluation to improve. Reflection enables self- correction without external feedback—the agent acts as its own critic. Common patterns include: output quality assessment, reasoning verification, and strategy evaluation.

What this is NOT

  • Not the same as external evaluation (reflection is self-assessment)
  • Not just generating multiple responses and picking one (that's sampling, not reflection)
  • Not Chain-of-Thought (CoT shows reasoning; reflection evaluates it)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

Prompting an LLM to critique or evaluate its own previous response, then using that critique to generate an improved response. Often implemented as a second LLM call that reviews the first.

Sources: Reflexion paper (Shinn et al., 2023), Self-Refine paper (2023), Constitutional AI (Anthropic)

cognitive-science

Metacognition—thinking about one's own thinking. The ability to monitor cognitive processes, assess their effectiveness, and adjust strategies accordingly.

Sources: Flavell: Metacognition (1979), Cognitive psychology literature

Examples

  • After generating code, the agent asks 'Does this handle edge cases?' and revises if not
  • An essay agent generates a draft, critiques it for argument strength, and rewrites weak sections
  • A research agent checks 'Did I cite sources for all claims?' and adds missing citations
  • Reflexion: agent attempts task, fails, reflects on why, tries again with adjusted strategy

Counterexamples

Things that might seem like Reflection but are not:

  • Human feedback on agent output (that's external evaluation)
  • Running unit tests on generated code (that's automated testing, not reflection)
  • Generating 5 responses and taking the longest one (no evaluation involved)

Relations

  • overlapsWith reasoning (Reflection is reasoning about one's own reasoning)
  • overlapsWith planning (Reflection can inform replanning)
  • overlapsWith agent-loop (Reflection can be a step within the agent loop)

See Also