ReAct: think, then act

Reasoning and actions interleaved, so each choice is made after seeing what the last one did.

Part of the Agents track on lAItest.

Ask a model to plan the whole job up front and it writes a confident plan for a world it has not looked at yet.

ReAct fixes that by forcing it to look between steps.

Reason and act, interleaved

ReAct is a prompting pattern, not a model feature. The paper is arXiv:2210.03629. Each step the model writes a short thought about the current situation, then one action, then reads the observation that comes back. Thought, action, observation, repeat. The thought is what lets the next action depend on what actually happened rather than on what was expected to happen.

A common misconception

Commonly believed: The reasoning text is decoration. It is an explanation printed next to the real work, and you could strip it out without changing anything.

Actually: In a ReAct-style loop the thought is part of the transcript, so the model reads its own earlier reasoning before it chooses again. It is an input, not a comment. Remove it and the behaviour changes.

Modern models do this without being asked

Frontier models now reason before answering by default, and the developer control is a qualitative effort level rather than a hand-written template. Older token-budget controls still exist alongside it on some models. So you rarely type Thought / Action / Observation yourself any more. The shape survived; it moved out of your prompt and into the model.

Why interleave reasoning with actions instead of planning the whole job first?

Answer: Because each observation can invalidate the plan, so the next choice has to be made after seeing it. A plan written before any tool has run is a guess about the environment. Interleaving lets each action depend on what actually came back. It usually costs more tokens rather than fewer — that is the trade you are choosing.

In one sentence

ReAct is a boring idea doing heavy lifting: decide one step at a time, after looking.