FRENES

Glossary · Agents and autonomy

ReAct

ReAct is a method for organizing an agent around alternating reasoning steps, actions and observations of their results. It lets a model adjust its approach using information gathered during a task, rather than requiring every action to be decided before the work begins.

ReAct: the explanation

The method connects intermediate reasoning with actions and the observations returned by tools, as described in the original [ReAct research](https://arxiv.org/abs/2210.03629). For a user, this means a search result can influence what the agent does next. You might see a record of searches, page visits or other operations. That record is useful for reviewing the work, particularly when the first source is incomplete or the system needs information that was absent from your initial request.

Consider a task where one document points to another source that needs checking. An agent organized this way can inspect that source before giving an answer. The practical benefit depends on whether those extra operations resolve the uncertainty. You should judge the evidence collected and the final outcome. A long activity history can make a process look thorough while still leaving the key question unanswered, especially if the agent keeps revisiting similar material without making progress.

ReAct does not itself grant permissions or define which tools are safe to use. The surrounding application still determines access, available operations and stopping conditions. A run can become unproductive if it keeps searching without finding enough evidence. Intermediate explanations should also be treated carefully. They are generated elements of the system’s operation, rather than a complete view into the model’s internal mechanisms, and they should be distinguished from actual observations returned by external tools.

A concrete example

You ask how to recover an item left on a train. The agent checks a general lost-property page and discovers that the procedure depends on the rail network. It asks which network you used, opens the relevant page and summarizes the details you will need to provide. The example shows a route adjusted through observations, without assuming the first page contains the complete answer.

Why it matters

Understanding the term can explain why an agent makes several tool calls during one request. When comparing systems, look at step limits, possible costs and error handling. A working method does not establish that the sources consulted are reliable. Its practical value depends on choosing useful actions, respecting access boundaries and checking whether the final result actually answers the task you gave it.

Common misconceptions

  • ReAct is not the name of a particular language model, because the same approach can be implemented with different models and tools whose capabilities, available information and limitations still affect the outcome of the task.
  • A sequence of intermediate explanations does not prove that an agent is reasoning correctly, because plausible steps can accompany a mistaken interpretation, weak evidence or a conclusion that goes beyond what the tools actually returned.

Related terms: agents and autonomy

AI agent
An AI agent is a system that uses a model to decide which steps to take toward a task.
autonomous agent
An autonomous agent is an AI agent that can pursue a goal and choose several actions without involving you at every step.
multi-agent system
A multi-agent system is a collection of agents that divide tasks or exchange information to pursue an objective.
planning (AI agents)
Planning in AI agents is the preparation of steps an agent intends to follow toward a goal.
agent memory
Agent memory is the set of mechanisms that lets an AI agent retain and retrieve information useful to its work.
human in the loop
Human in the loop is an arrangement in which a person participates at selected stages of an AI system’s work.
tool calling
Tool calling is a mechanism through which a model asks software to run a specific function, such as searching documents or checking a calendar.
computer use
Computer use is a capability that lets an AI agent interact with a graphical interface by clicking, typing or scrolling.

Where this site covers it

The term appears on 53 pages of this site. A few of them:

Same family

Go further

Frequently asked questions

Is this the React library?

No. In discussions of agents, ReAct refers to an approach connecting reasoning and action. React is also the name of a library for building user interfaces. The surrounding context tells you which meaning applies.

Must every step be visible?

An interface does not necessarily show every intermediate step. To assess the work, focus on completed actions, tool results and verifiable evidence, rather than treating a detailed narrative as proof that the outcome is reliable.

When should the loop stop?

The system can stop when the objective is met, clarification is needed or a configured limit is reached. These conditions should prevent unproductive repetition and allow the agent to report that work remains incomplete.

Neutrality. No affiliate links. No vendor pays to be reviewed, rated or ranked. Reviews are directed by François Chollet, neutral and factual, and every data point is dated and sourced. How we score