ReAct (Reasoning and Acting)

A pattern in which a model interleaves reasoning traces, tool actions, and observations returned by the environment. Each observation updates the next step, allowing the system to gather information, revise a plan, or stop when the objective is met.

ReAct structures an execution loop; it does not make proposed actions trustworthy. The application still needs to restrict tools, validate arguments, control repetition, define budgets, and record the trace. In sensitive settings, textual reasoning may also contain private data or rationales that are not faithful to the internal mechanism.