What's the difference between tool calling and the ReAct pattern?

Tool calling is the model-API mechanism: the model emits a structured call naming a function and its arguments. ReAct is an older prompting pattern, predating native tool-calling APIs, that interleaves a reasoning step with each action in a loop instead of just taking the action.

Yao et al.’s 2022 paper describes generating “reasoning traces and task-specific actions in an interleaved manner”: the reasoning step helps the model track and update its own plan and catch exceptions, while the action step lets it reach an external source, like a knowledge base or a tool, for information it doesn’t already have. Submitted October 6, 2022, ReAct predates OpenAI’s function-calling API, which launched June 13, 2023, by about eight months, so its original implementations had no structured call to emit: the model wrote a “Thought” and an “Action” as free text, and the surrounding code parsed that text for the tool name and its input.

Most ReAct implementations today don’t parse free text at all. LangChain’s own agent builder runs the ReAct loop on structured tool calls instead, so the reasoning step and the tool call are separate things in the same request, not one being how the other gets read out.