AIToday
Large Language ModelsZenn AI/MLPublished: Nov 21, 2026, 16:00 JST

ReAct: Google Research's 2022 loop behind AI agents

Send on LINE
ReAct: Google Research's 2022 loop behind AI agents

3 Key Points

  1. What happened

    A Zenn explainer lays out ReAct, the 2022 Google Research pattern in which an LLM alternates Reasoning and Acting through a Thought → Action → Observation loop until a finish action.

  2. Why it matters

    Splitting a task into explicit thoughts, tool calls and observations lets the model use outside information and show why it chose each step, which the explainer says makes agent behavior easier to trace.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The explainer situates ReAct against earlier approaches that handled only one half of the problem. Methods focused on reasoning alone could not reach outside information, while methods focused on tool use alone gave no account of why an action was chosen. ReAct combines the two so that each thought guides the next action and each observation enriches the next thought.

The pattern is presented as an early-2022 proposal from Google Research, and the explainer argues it became a foundation for modern AI agents for three reasons: the mutual reinforcement of reasoning and acting, its resemblance to how people alternate thinking and checking, and the traceability that comes from writing thoughts out. It also lays out where sibling patterns fit: plan-first execution for predictable steps, self-critique and retry for trial-and-error tasks, and branching searches over multiple reasoning paths for wide search spaces. The caveat is that the added reasoning is not free — on simple tasks, skipping it can be quicker.

FAQ
What are the three steps in the ReAct loop?
Thought, Action and Observation repeat until the model issues a finish action. Thought states the situation and next plan, Action runs a tool with arguments, and Observation feeds the result back into the next thought.
Does ReAct always beat other approaches?
No. The explainer says ReAct is not the best choice for every task, and on simple tasks skipping the reasoning steps can be faster.
Why is ReAct easier to debug?
Because the model writes out its Thought — why it chose an action — as explicit text, humans can follow and verify the agent's decision process.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleContra Pace: Ben Pace's one-graph case for LessWrong's writer decline