Reflexion

Reflexion is an agent pattern introduced by Shinn et al. in a 2023 paper that adds explicit self-reflection and learning from past attempts, enabling agents to improve performance on repeated trials of the same task. The technique works by having the agent attempt a task, evaluate its own performance against criteria, store reflective summaries of what went wrong, and use those reflections to inform subsequent attempts. Reflexion was shown to substantially improve LLM performance on coding tasks (HumanEval) and decision-making tasks (ALFWorld) compared to single-shot baselines. The pattern is particularly valuable for agentic workflows where the agent has multiple chances to succeed — code generation with test feedback, multi-step research with intermediate verification, iterative refinement workflows. Reflexion has influenced many subsequent agent designs and is supported by frameworks like LangGraph and AutoGen. AI governance teams document reflexive agent behavior in AI compliance lineage because the agent's stored memory and reflection history are part of the system's learned state. The original Reflexion code is hosted at github.com/noahshinn/reflexion.

Reflexion-enabled agents with Centralpoint: Centralpoint supports Reflexion-style self-improving agents with any LLM in a model-agnostic stack, with full reflection-history audit logs. Tokens are metered per skill and iteration, prompts stay local, and self-improving chatbots deploy through one line of JavaScript on any portal.


Related Keywords:
Reflexion,,