← Reference · Nestor G Pestelos Jr · Print this page

Systems Architecture · Artificial Intelligence

Reflexion (AI Agents)

Reference entry · last updated 20260910

Reflexion is a framework in which a language-model agent reflects on feedback and stores verbal lessons in episodic memory to inform later attempts. This memory holds records of earlier attempts. Shinn and colleagues introduced it as verbal reinforcement learning.[1]

First principles and definitions

Reflexion changes the information supplied to an agent across attempts without updating model weights. Its feedback may be a score or free-form text, supplied by external checks or generated internally.[1]

Mechanism

An actor attempts a task. An evaluator assesses the result. A self-reflection component produces text retained in memory for a subsequent attempt. The actor can use ReAct within an attempt.[1]

Illustrative sequence: a code candidate fails a boundary-case test. A reflection records the missed case. On the next attempt, the actor receives that note alongside the task. A new test run determines whether the revision fixes the failure.

This example uses a test as feedback; the framework does not require every task to have a deterministic verifier. The mechanism also does not imply that the agent rewrites its tools.

Evidence and limits

The paper evaluates decision-making, reasoning, and programming tasks. A mistaken reflection can misdirect later attempts; retained text is not verified knowledge.[1]

An agent harness still manages execution.[2] An agentic loop using Reflexion needs its own stopping policy and task-appropriate evaluation.

See also

• Agentic Loops
• ReAct (AI Agents)
• Context Engineering
• LLM Evaluations

References

1. Shinn, Noah, et al. “Reflexion: Language Agents with Verbal Reinforcement Learning.” NeurIPS, 2023. Full text.
2. Trivedy, Vivek. “The Anatomy of an Agent Harness.” LangChain, 2026.