Reflexion
Agent pattern where the LLM verbally reflects on failed attempts, stores the reflection in memory, and tries again — "verbal reinforcement learning" without gradient updates.
Shinn et al. (Northeastern, 2023). Loop: actor produces an action → evaluator gives feedback (right/wrong + explanation) → reflector writes a lesson in text → actor reads previous lessons before the next attempt.
Showed sharp improvements on HumanEval (code) and WebShop (web navigation) using only prompt engineering. Inspired a whole class of "self-correction" techniques that are now default in agentic pipelines (Cursor, Aider, Devin).