ReAct prompting combines reasoning and acting in a repeated pattern: the model considers the task, selects an action, observes the tool result and continues. This can help a system use external information rather than trying to answer everything from its initial context. Implementations vary in how they represent or expose reasoning.
A research agent using this pattern might reason that it needs the current price, call a search tool, read the result, then decide whether another lookup is needed before answering. Modern model APIs often provide built-in tool calling, which can handle part of this pattern more reliably than parsing free text. ReAct remains a useful way to think about how an agent loop alternates between thinking and acting.
Tool use must still be controlled by application code. Validate arguments, restrict permissions and stop repeated or irrelevant actions. ReAct is a design pattern, not a guarantee of correctness. Related: AI Agent, Tool Calling.