Reflexion
Also known as: reflexive agent
Reflexion is a method where an agent evaluates its own attempt, writes a natural-language self-critique, and uses that reflection to improve on the next try.
- After a failed step, the model asks: what went wrong?
- The critique is stored and fed back on the next attempt.
- A form of in-session learning without weight updates.
The pattern's power: it turns failure into *usable* signal. An agent that can articulate why it failed does far better on the retry than one that just tries again blindly.