Most tool-using agents you'll encounter are built on a pattern called ReAct — short for Reasoning + Acting. It sounds academic, but the idea is simple and once you see it, agent behavior becomes readable. Here is the model.
The Core Loop: Thought, Action, Observation
ReAct interleaves three things in a loop. The model produces a thought (reasoning about what to do), takes an action (calling a tool), and receives an observation (the tool's result). Then it reasons again, using what it just observed:
Thought: I need the user's order status. I should look it up.
Action: get_order(id="1234")
Observation: Status = "shipped", ETA = "Aug 9"
Thought: I have what I need. I can answer now.
Answer: Your order shipped and arrives Aug 9.
Why Interleaving Matters
The power comes from mixing reasoning with acting instead of doing all the thinking up front. Because the model reasons after each observation, it can adapt — if a lookup fails or returns a surprise, the next thought accounts for it. A plan made entirely in advance can't react to what it learns; ReAct can.
Reasoning Makes Actions Better
Explicitly generating a thought before each action isn't decoration — it measurably improves the model's tool choices. Writing out "I need X, so I should call Y" grounds the next action in a stated reason, the same way chain-of-thought improves plain reasoning tasks.
How It Ends
The loop continues — thought, action, observation, thought — until the model decides it has enough to answer. At that point it emits a final answer instead of another action, and the loop stops:
Thought: I now have all the information needed.
Answer: <final response to the user>
Summary
ReAct is the pattern of interleaving reasoning, acting, and observing in a loop: the agent thinks, calls a tool, sees the result, and thinks again — adapting as it goes rather than planning everything up front. The explicit "thought" before each action makes its tool use sharper and its behavior traceable. Understand ReAct and you understand the backbone of nearly every agent framework in use today.