Skip to main content

Reflection

After generating output, the agent explicitly evaluates its own work — identifying errors, weaknesses, and missing elements — then uses that self-critique to produce an improved version. The agent acts as its own reviewer. This can be single-pass (one reflection) or iterative (reflect-revise loop).

Structure​

The agent generates, critiques, and revises in a self-contained loop. Unlike Evaluator-Optimizer (an orchestration pattern with two separate roles), reflection happens within a single agent.

How It Works​

  1. Generate — produce an initial output
  2. Critique — evaluate the output against quality criteria (prompted self-review)
  3. Identify issues — list specific problems, gaps, or improvements
  4. Revise — incorporate the critique and produce an improved version
  5. Repeat or finish — loop until satisfactory or max iterations reached

The critique prompt matters. Effective patterns:

  • "What's wrong with this output? List specific issues."
  • "Score this output 1-10 on accuracy, completeness, and clarity. Explain each score."
  • "If you were reviewing this, what feedback would you give?"

Key Characteristics​

  • Self-improving — output quality increases with each iteration
  • Single agent — no separate evaluator needed (unlike Evaluator-Optimizer)
  • Diminishing returns — most improvement happens in the first 1-2 reflections
  • Self-blindness risk — the agent may not catch its own systematic errors
  • Higher cost — each reflection cycle is an additional LLM call

When to Use​

  • Writing and editing tasks where revision improves quality
  • Code generation — generate, then review for bugs and edge cases
  • The agent makes systematic errors that it can catch with a targeted critique prompt
  • You want quality improvement without the complexity of a separate evaluator agent
  • Tasks where a human would naturally review their own work before submitting