What is ReAct prompting?

ReAct prompting — short for Reasoning + Acting — is a technique where you get a model to interleave its thinking with real actions instead of just reasoning in its head. It works in a repeating loop: the model writes a Thought about what to do next, takes an Action (like a tool or search call), reads the Observation that comes back, and then thinks again. That loop is the foundation under most modern AI agents.

How does the Thought-Action-Observation loop work?

Each turn, the model produces a short reasoning trace explaining its plan, then commits to a concrete action, then waits for the result before continuing. The observation grounds its next thought in real data rather than a guess, so it can correct course mid-task.

1. Thought — the model reasons out loud about what it needs and why.
2. Action — it calls a tool: search, a calculator, a database query, an API.
3. Observation — the tool's output is fed back into the context.
4. Repeat — it keeps looping until it has enough to give a final answer.

How is ReAct different from chain-of-thought?

Plain chain-of-thought prompting asks the model to reason step by step, but all of that reasoning happens inside its own head with no access to the outside world. If a fact is wrong or out of date, chain-of-thought will confidently reason from the bad fact and get a wrong answer.

ReAct adds the acting half. Between reasoning steps, the model can go check reality — look something up, run a calculation, read a file — and fold what it learns back into the next thought. That grounding cuts down on made-up facts and lets the model handle tasks that need fresh or external information.

Why does ReAct matter for building agents?

The Thought-Action-Observation cycle is essentially the agent loop. When people talk about an LLM agent that plans, uses tools, and adjusts, ReAct is the prompting pattern making that possible under the hood. A few practical notes:

• It's a prompting strategy, not a separate model — you can apply it to most capable LLMs.
• It pairs naturally with tool use and function calling, which supply the actions.
• Long loops can drift or repeat, so real systems add step limits and guardrails.

In short, ReAct is what turns a model that only talks into one that can think, act, and check its work in a loop.