ReAct Pattern
Thinking and doing are usually separated in software: work out the answer, then act on it. The ReAct pattern refuses that separation. The agent reasons one short step, takes one action, looks at what came back, and reasons again — so every thought is informed by the last observation rather than by a guess about it.
The loop, concretely
Each turn produces a small, legible trace:
- Thought — what do I know, and what should I try next?
- Action — one tool call, with arguments.
- Observation — whatever the tool returned, appended to the context.
Then it repeats until a stopping condition fires. The name contracts "reason and act", from the 2022 paper that introduced it, and the contribution was showing that interleaving the two beats either alone: pure chain-of-thought reasons itself into confident fiction, while pure tool-calling has no mechanism for deciding what to call next.
Why it became the default
Grounding. Each observation pulls real evidence into the context, so the reasoning is corrected by the world every turn rather than at the end. That makes ReAct robust where plans are unreliable — exploratory research, debugging, anything where step three depends on what step two actually returned. The costs are equally characteristic: the context grows with every observation until it must be compacted, and a loop with no step cap will cheerfully retry a failing tool until the budget is gone. Almost every modern agent runtime, from LangGraph to vendor SDKs, ships a ReAct-shaped loop as its starting point.
The ReAct pattern interleaves reasoning and action in a single loop: the agent emits a short thought, makes one tool call, reads the observation, and reasons again from it. Introduced in 2022 as "reason and act", it grounds each decision in real evidence instead of a plan written in advance, which is why it underpins most agent runtimes. Its costs are a context that grows with every observation and a loop that needs an explicit step cap.
Picture someone finding their way in an unfamiliar city without a map. They do not plot the whole route in advance. They look, take one turn, see where that got them, and decide again.
That is the ReAct pattern: the way most AI agents work. Think a little, do one thing, look at the result, think again. The name is short for "reason and act".
It works because the agent keeps checking against reality instead of trusting its own guesses. The downsides are familiar to anyone wandering a city: you can go round in circles, and without someone saying "enough", you can keep trying forever.