ReAct

Gold definitionUpdated Apr 2, 2026

Definition

ReAct (Reasoning and Acting) is an LLM-based paradigm that enables agents to interleave reasoning traces with task-specific actions. This allows models to dynamically plan, execute tools, and observe results, leading to more robust and reliable problem-solving in complex environments.

At a glance

Executive summary

ReAct is a method for AI agents to think and act by having a large language model (LLM) generate thoughts and then execute tools, observing the results. This iterative process helps the agent solve complex problems more reliably, like automating software testing or creating new data features.

TL;DR

ReAct lets AI models reason through problems step-by-step and use external tools, making them smarter and more reliable at complex tasks.

Key points

Interleaves explicit reasoning (thoughts) with task-specific actions (tool execution) for dynamic problem-solving.
Addresses issues like hallucination and poor reasoning in LLM-based agents, enabling robust automation of complex workflows.
Used by researchers and ML engineers developing advanced AI agents for tasks like automated software testing, feature engineering, and planning.
Improves upon LLM-only baselines by integrating tool use and iterative refinement, leading to more grounded and reliable outcomes.
A foundational paradigm in the rapidly evolving field of agentic AI systems, focusing on enhancing LLM capabilities through external interaction and self-correction.

Use cases

Automated Software Testing: Generating specialized fuzz harnesses for Java libraries by decomposing the workflow and querying documentation on demand.
Automated Feature Engineering: Autonomously exploring, generating, and refining features for tabular data in machine learning models (e.g., FAMOSE framework).
Complex Natural Language Planning: Serving as a baseline for LLM-based planning systems to handle multi-step tasks and achieve specific goals.
Code Generation and Repair: Agents reasoning about code, generating snippets, and using compilers/linters as tools to repair errors.
Data Analysis and Manipulation: Allowing LLM agents to interact with databases, APIs, and data processing tools to perform complex data analysis tasks.

Also known as

ReAct paradigm, ReAct agents, Reasoning and Acting