# What is ReAct Pattern?

Also called ReAct, reason and act.

ReAct, short for reasoning and acting, is an agent pattern in which the model alternates between writing out a thought about what to do next and taking an action such as a tool call, then observing the result before thinking again. Introduced in a 2022 research paper, it became the default shape for tool-using agents.

The pattern produces a trace of alternating entries: thought, action, observation, thought, action, observation. The thought is plain reasoning that is never executed. The action is a structured request the runtime can run. The observation is whatever came back, including errors. Because each observation lands before the next thought, the model can react to reality rather than committing to a plan written up front.

Interleaving matters because reasoning alone drifts from facts and acting alone has no strategy. Writing a thought before each action forces the model to state why it is doing something, which improves tool selection. Feeding the observation back lets it notice a failed search or an empty result and change approach, which a pre-planned sequence cannot do.

ReAct is not magic and has known weaknesses. Traces grow long, consuming context and cost. Agents can loop, repeating a failing action with slight variations. The written thought is a useful artifact but should not be read as a faithful account of the model's internal computation. Step limits, loop detection, and forced summarization are standard mitigations.

Modern implementations rarely use the original text format. Structured tool calling supplies the action slot natively, and reasoning models produce the thought internally, so the visible trace is often just calls and results. The pattern survives as the underlying shape. It is closely related to the agent loop, chain of thought, and tool use, and most agent frameworks implement it by default.

## Key points

- Alternates thought, action, and observation in a repeating cycle.
- Observations arrive before the next decision, allowing course correction.
- Published in 2022 and now the default agent shape.
- Prone to long traces and repeated failing actions without limits.

## In practice

Asked for a company's latest funding round, an agent thinks that it needs recent news, searches, and observes three results, none newer than last year. It thinks the query was too broad, searches again with the current year added, and observes a press release from March. It then thinks it has enough and answers. Three thoughts, two actions, two observations.

## Related terms

- [Agent Loop](/en/glossary/agent-loop)
- [Chain of Thought](/en/glossary/chain-of-thought)
- [Tool Use](/en/glossary/tool-use)
- [Agent Trajectory](/en/glossary/agent-trajectory)
- [AI Agent](/en/glossary/ai-agent)

[Back to the AI Glossary](/en/glossary)