Self-reflection
An agent capability where the model generates an explicit reflection on its own reasoning or outputs — used to improve subsequent steps or detect errors before they propagate.
Self-reflection is a softer version of self-correction. Rather than directly fixing an error, the agent generates a reflection — "here is what I tried, what worked, what failed, what I would do differently" — and uses the reflection as context for the next attempt or step.
The pattern shows up in many places: chain-of-thought reasoning has implicit reflection; the Reflexion pattern uses explicit reflection between attempts; multi-agent systems often have a dedicated "reflector" or "critic" agent role.
In 2026, self-reflection is most useful in longer-horizon agent runs where mistakes compound. After every 5–10 actions, having the agent reflect on progress and adjust strategy materially improves outcomes on complex tasks.
Frequently asked
When should I add self-reflection to my agent?+
For agents running 10+ steps. Self-reflection between major milestones catches drift and lets the agent course-correct. For short tasks (1–3 steps), the overhead usually outweighs the benefit.
Self-reflection vs chain-of-thought?+
CoT generates reasoning before an answer. Self-reflection generates reasoning about a past attempt. CoT improves single-shot quality; self-reflection improves multi-step recovery from errors.