How does ReAct differ from pure chain-of-thought, and what does interleaving reasoning with acting provide?

**Chain-of-thought** asks the model to think out loud before answering. The model writes a reasoning trace, step by step, but every word in that trace comes from its own prediction. No real data enters after the first prompt. ReAct changes the shape of the trace. Instead of pure reasoning, it interleaves three kinds of steps. A Thought is the model reasoning about what to do. An Action is a tool call, such as a search or a calculation. An Observation is the real result that tool returned. The loop runs Thought, then Action, then Observation, then back to Thought, until the task is done. The crucial part is that each Observation is a genuine fact from the world, not something the model made up. This is what interleaving buys you. In a long chain-of-thought trace, one early mistake silently corrupts everything after it, because nothing ever checks the work. In ReAct, the next Thought reads a real Observation, so the reasoning keeps getting pulled back to reality instead of drifting further into a guess.

**Chain-of-thought** and ReAct both produce a step by step trace, but they differ in what feeds each step. Chain-of-thought reasons entirely in the model's imagination. After the prompt, no real data enters, so every step is conditioned only on prior generated tokens. ReAct interleaves three token types in one trace: Thought, Action, and Observation. A Thought is reasoning. An Action is a tool call. An Observation is the actual result that tool returned. The cycle repeats until the answer is found. The payoff of interleaving is grounding. After each Action, the next Thought reads a real Observation, not a hallucinated continuation. This anchors reasoning to facts and prevents the hallucination drift that accumulates over long pure reasoning traces. It also enables early stopping. If an Observation already answers the question, the agent terminates instead of generating more ungrounded steps. The cost is one extra LLM call per step. Plan then act runs the model once, but ReAct pays a round trip for every Thought-Action-Observation cycle. You trade latency and tokens for the ability to recover when the world disagrees with the plan.