Reflexion: Self-Critique, Feedback, and Iterative Improvement in Agentic AI
Reflexion: Self-Critique, Feedback, and Iterative Improvement
Reflexion in one sentence
Reflexion is a pattern where the agent attempts a solution, evaluates it, writes a short “lesson”, and uses that lesson to do better on the next attempt.
Why reflection works
LLMs are surprisingly good at catching their own mistakes when you ask them to review with a different lens: “Check for missing constraints”, “Verify numbers”, “Look for contradictions”.
A practical 3-pass structure
- Draft: produce the best answer/plan
- Critique: find issues (factual gaps, constraint violations)
- Revise: fix issues and produce final output
Where to store reflections
Store reflections as compact bullets in a long-term memory store (vector DB or key-value). You don’t want to store entire conversations—store “rules of thumb” the agent can reuse.
Anti-pattern: reflection loops forever
Set a max number of revisions. In practice, 1–2 rounds of critique is enough for most tasks.
What reflexion is NOT
Reflexion is not fine-tuning. It’s runtime self-improvement via structured feedback.
