Agent Control Patterns — Part 3: Reflexion — When Review Triggers Research

2026-03-03 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, long

Summary

The Reflexion pattern extends the Reflection technique for large language models by integrating external research to improve factual accuracy, not just clarity or reasoning. While Reflection improves answers using only the model's internal knowledge, Reflexion introduces a structured workflow: Draft → Self-Critique → Generate Search Queries → Run Tools → Revise Using Evidence → Stop. This process is enabled by a structured `DraftOutput` Pydantic model, which forces the model to produce an initial answer, a self-critique identifying gaps, and specific search queries. These queries are then executed by an external tool, like a web search API, and the retrieved evidence is used to inform an evidence-based revision, resulting in a `FinalOutput` that includes references. This multi-step, controlled approach, implemented via separate nodes in a graph, allows the system to correct outdated or factually incorrect information, especially for time-sensitive questions or tasks requiring verified claims.

Key takeaway

For AI Engineers building systems that require high factual accuracy or deal with time-sensitive information, adopting the Reflexion pattern is crucial. This approach allows your models to move beyond internal reasoning by programmatically triggering external research and incorporating new evidence. Implement structured output formats for drafts and revisions, and integrate reliable search tools to ensure claims are verifiable and up-to-date, thereby enhancing the trustworthiness of your AI applications.

Key insights

Reflexion enhances LLM factual accuracy by integrating external research into a structured self-correction loop.

Principles

• Separate generation from evaluation.
• Structure critique to trigger research.
• Evidence-based revision improves accuracy.

Method

The Reflexion workflow involves drafting, self-critiquing to generate search queries, executing queries via external tools, and then revising the answer using the retrieved evidence, often with iteration limits.

In practice

• Use Pydantic models for structured LLM outputs.
• Integrate web search APIs for external evidence.
• Implement iteration limits to prevent loops.

Topics

• Reflexion AI Pattern
• External Knowledge Integration
• Evidence-Based Revision
• AI Workflow Orchestration
• ReAct Pattern

Best for: Machine Learning Engineer, AI Engineer, Research Scientist

Related on AIssential

• Agent Control Patterns — Part 2: Reflection — A Simple Way to Improve Answer Quality — Separating generation from evaluation through reflection improves AI answer quality without model retraining.
• ForeSci: Evaluating LLM Agents for Forward-Looking AI Research Judgment — ForeSci benchmarks LLM agents' forward-looking research judgment, revealing evidence-decision decoupling despite explicit evidence organization.
• Prompt Engineering for Agentic AI — Agentic AI prompting requires architectural context engineering, not just better phrasing, to ensure reliable multi-step behavior.
• I Built an Agentic AI Report Writer That Reviews and Corrects Its Own Sections — An agentic AI system automates structured report writing from diverse enterprise sources using self-correction and dependency management.
• MARCH: Multi-Agent Reinforced Self-Check for LLM Hallucination — MARCH uses multi-agent reinforcement learning and information asymmetry to reduce LLM hallucination in RAG systems.
• Evaluating LLMs as Human Surrogates in Controlled Experiments — LLMs can reproduce directional behavioral effects but often miscalibrate effect magnitudes compared to human data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.