What is the ReAct pattern?
Answer
ReAct stands for Reasoning and Acting. The model alternates between producing a thought about what to do next and taking an action, then observes the result before the next cycle.
The cycle is thought, action, observation, repeated. The value is that reasoning is conditioned on real observations rather than on assumptions, so the model can correct course when a result contradicts what it expected.
The alternative it was contrasted against is plan and execute, where the model produces a complete plan first and then runs it. That is more predictable and easier to review before anything happens, but it handles surprises badly, because the plan was written without knowing what step two would return.
In practice most modern coding agents are effectively ReAct: they decide the next action based on what just happened, rather than committing to a full plan in advance.
Common wrong answer
Treating ReAct as a specific library or framework feature you enable. It is a prompting pattern, and its structure is now largely built into how tool calling models behave by default.
A more subtle mistake is claiming ReAct is strictly better. For tasks that are well understood and where you want a human to approve the whole approach before anything runs, plan and execute is the safer design. The choice depends on how much you need reviewability versus adaptability.
With models that do extended internal reasoning, the explicit thought step is often redundant, since the model already reasons before emitting a tool call.
Likely follow-ups
- How does it compare to plan and execute?
- Is it still relevant with reasoning models?