ReAct

advancedwidespread

ReAct is an operational pattern where the AI agent alternates between thinking and acting: it weighs options, uses a tool, checks the result, then weighs again.

In plain terms

Imagine a plumber coming to fix a dripping faucet. They do not figure out the entire repair in advance but check what might be wrong, tighten a nut, turn on the water, and decide the next step based on what they see. ReAct gives this same rhythm to the AI agent: thought, action, observation, then repeat.

The name is a combination of two English words: reasoning and acting. The essence of the pattern is that the language model does not answer in a single attempt but uses tools in the process, such as searching the web, opening a file, or running a calculation, and incorporates the obtained result into its further line of thought.

This differs from a plain chatbot that works only from what it was trained on. An agent built on ReAct can look things up and notice along the way if it has set off in the wrong direction. That is why it became one of the basic patterns for the agents known today.

How it works

It receives the task and a list of tools it can use (for example, a search engine, calculator, or file reader).
Thought: the model describes to itself what it knows so far and what the next step should be.
Action: selects a tool and specifies what to request from it, such as a search query.
Observation: the tool's output feeds back into the model and becomes the basis for the next thought.
The cycle repeats until the agent deems the answer found or it reaches the set step limit.

An everyday example

Ask an AI assistant which train gets you to Pécs by noon on Saturday and how much the ticket costs. On screen, you see it first check the schedule, then ticket prices, while displaying what it is doing, before finally answering. This searching, step-by-step operation is the ReAct pattern.

Who is affected

developerIf you build an agent, you will almost certainly use this cycle or one of its derivatives, even if the framework hides it from you. → Code

employeeOffice AI assistants thus perform multi-step tasks, so it is useful to know why they take longer and where errors might occur. → Work & HR

entrepreneur, executiveStep-by-step operation requires more model calls, meaning higher costs and longer wait times, but in return, the agent's work is traceable.

everyoneWhen a chat program displays 'searching' or 'opening file', meaning it uses a tool while thinking, such a cycle is likely running in the background.

A common misconception

MythReAct is a specific program or product that can be downloaded, and it is related to the web development tool called React.

In realityReAct is an operating pattern that can be implemented with many different models and frameworks. The similarly named React is a programming library for building web interfaces; the two have nothing to do with each other.

Limits and risks

The described reasoning is not a guarantee of correctness: an agent can argue convincingly for an incorrect step.
Each round is a new model call, so long tasks can be slow and expensive, and the agent sometimes loops through the same steps repeatedly.
The result is only as good as the tools: if the search engine returns poor or misleading results, the agent builds its next steps on them.
If an agent can perform real actions (sending emails, making payments, deleting files), a faulty step has real consequences, necessitating step limits and human approval.

Similar, but not the same

Chain of thought – In chain-of-thought, the model reasons internally step by step; ReAct adds tool usage and processing of returned results to this.
Tool use – Tool use is the ability of a model to call external functions, while ReAct is the loop that alternates tool calls with reasoning to turn them into task solving.
Planner-Executor pattern – There, the agent first creates a multi-step plan and executes it, adjusting only occasionally. ReAct, however, does not plan ahead but re-decides on continuation after each step.

Related concepts

Broader

AI-agent, Agent patterns, prompting techniques

Narrower

Thought–Action–Observation cycle

On the edge

Chain of thought, Tool use, Function Calling, Planner-Executor pattern, Agentciklus

A brief history

The name and method ReAct were introduced by Shunyu Yao and co-authors, researchers from Princeton University and Google, in a paper published in 2022, which was presented at the ICLR conference in 2023. The pattern quickly entered open-source agent frameworks, becoming one of the foundational concepts of agent building.

Where things stand

By the end of 2026, alternating between thinking and acting is the basic operation of most AI agents, but it now mostly relies on built-in tool calling in models rather than the original text-based prompt format. The name ReAct itself survives mainly in professional materials and framework documentation.

Status: 2026 November

In the news

Appears in 24 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Reactors under construction rise to 79 worldwide as AI demand grows; SMRs expected to spread fastOct 28, 2026 · Ipar
Nvidia and PreussenElektra assess decommissioned German nuclear sites for AI data centresOct 27, 2026 · Ipar
Amazon funds 12 small modular reactors at a former coal plant site in IndianaOct 27, 2026 · Ipar
US restarts shut-down reactors and expands operating ones to power data centersOct 26, 2026 · Ipar
Etsy restricts AI-generated content to curb seller slopOct 23, 2026 · Marketing
US prepares to restart three shut-down nuclear reactors for AI data centersOct 20, 2026 · Ipar
Amazon plans nearly 1 GW small modular reactor project in IndianaOct 19, 2026 · Ipar
Nuclear power for data centers: big tech's SMR deals and the existing fleetOct 1, 2026 · Ipar

Also known as

ReAct · ReAct agent · ReAct-minta · ReAct-ciklus · ReAct agent · reasoning and acting · ReAct prompting · ReAct loop

Sources

Yao et al.: ReAct: Synergizing Reasoning and Acting in Language Models (arXiv, 2022)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.