ReAct

Harness · Reason + Act

A way for an agent to work in which the model alternates reasoning and acting: it decides what to do, calls a tool, reads the result and reasons again. Reasoning alone, the model makes the answer up; acting alone, it wanders without a plan; alternating, it gets to the right answer. Described in a 2022 paper; the agent loop of today’s coding agents rests on it.