← Reference · Nestor G Pestelos Jr · Print this page

Systems Architecture · Artificial Intelligence

ReAct (AI Agents)

Reference entry · last updated 20260910

ReAct is a method that interleaves language-model reasoning with actions and observations from an environment. Yao and colleagues introduced it for tasks requiring both deliberation and interaction.[1]

First principles and definitions

An action can obtain evidence or change an environment. Its observation becomes input for later decisions in an agentic loop. ReAct combines these interactions with generated reasoning traces.[1]

Mechanism

The original method uses example trajectories to elicit reasoning and actions. Reasoning can be sparse; the method does not require a thought before every action.[1]

Illustrative sequence: an agent searches a paper title, reads a candidate record, notices a title mismatch, and changes the query. Each observation is available when choosing what to do next.

The harness executes requested tools. Approval checks and execution limits are separate runtime decisions.[2] ReAct itself does not require human approval for each action.

Evidence and limits

The paper evaluates question answering, fact verification, and interactive tasks using HotpotQA, FEVER, ALFWorld, and WebShop. Those results concern the tested settings.[1]

ReAct can still choose an unhelpful action or draw a wrong conclusion. Reflexion can add reflection across attempts.

See also

References

  1. Yao, Shunyu, et al. “ReAct: Synergizing Reasoning and Acting in Language Models.” ICLR, 2023. Author project page.
  2. Trivedy, Vivek. “The Anatomy of an Agent Harness.” LangChain, 2026.