Why it matters

Reflexion improves task completion. Understanding shapes advanced agent.

Advertisement

The architecture

Try task.

Evaluate: success or failure.

Reflect: what went wrong.

Retry with reflection.

Reflexion patternAttempt taskinitial approachEvaluatesuccess or failReflect + retrywith lessonsBounded retries; reflection stored per task or across
Reflexion loop.
Advertisement

How it works end to end

Evaluator: LLM or ground truth.

Reflection: LLM analyzes trajectory.

Retry: with reflection in context.

Memory: across episodes optional.