ReAct
A pattern where the model alternates between thinking out loud and taking an action, each step informing the next.
Think of it like
A detective who mutters a theory, checks one clue, updates the theory, checks another — reasoning and legwork taking turns.
Example
“I need the population — let me search.” It searches, reads the result, then reasons: “Now I can compute the density,” and calls the calculator.
How it actually works
ReAct interleaves a reasoning trace with tool actions and their observations, so the model’s next thought is grounded in what it just saw rather than pure speculation. It made early agents far more reliable than plan-everything-upfront approaches, because the model can course-correct after each real result. The reasoning steps also make its behavior legible when something goes wrong.
For product teams
The recipe behind most working agents — think a little, act a little, repeat.
For engineers
An agent loop interleaving chain-of-thought tokens with tool actions and observations fed back into context.
Related
- Chain of thought — The think-out-loud half of the pattern.
- Tool Use — The take-an-action half.
- Observation — The result an action returns before the next thought.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome