Reflection

Reflection: after the agent produces an answer or a plan, it's prompted to critique itself. "Is this correct? Is anything missing? What would I change?" The revised answer is often meaningfully better than the first.

The pattern

1. Agent produces initial output (answer, plan, code, whatever)
2. Reflection step: another LLM call asks "review this output, find issues, suggest improvements"
3. Revision step: agent produces a new version incorporating the reflection
4. Optional: loop reflection-revision until satisfied or max iterations

When reflection wins

• Code generation (catches bugs before shipping)
• Complex reasoning (catches logic errors)
• Writing tasks (catches tone, factual, structural issues)
• Plans (identifies missed steps or bad assumptions)

When it doesn't

• Simple tasks where the first answer is fine
• Cost-sensitive high-volume workflows
• Latency-critical apps
• When the model keeps producing the same "reflection" (it's not learning from itself)

Self vs separate critic

Two approaches:

• Self-reflection: same model critiques its own output. Cheaper, faster. Limited by the model's self-awareness.
• Separate critic: a different model (or different prompt persona) critiques. More expensive. Often produces meaningfully different feedback.

For high-stakes outputs (production code, published content), separate critic is worth the cost.

The empirical gain

On standard benchmarks, reflection improves output quality by 5-20% over no reflection. Bigger gains on complex reasoning tasks, smaller on factual retrieval.

Infinite reflection loops

Without a stop condition, reflection can run forever. Cap at 2-3 iterations and accept diminishing returns.