ReAct (Reasoning + Acting)

ReAct — short for Reasoning + Acting — is a pattern for building agents that weaves a model's thinking together with the actions it takes. Instead of producing an answer in one breath, the model emits a short Thought about what it needs next, an Action that calls a tool, and then receives an Observation: the actual result of that call. It reads the observation, thinks again, and the cycle repeats until the model decides it has enough to finish.

The reason this works so well is grounding. A model reasoning purely from memory can drift into a confident but wrong conclusion. By forcing every reasoning step to be checked against a real result, ReAct keeps the trajectory anchored to facts and lets the model notice when a search came back empty or a value looks off — and re-plan instead of barreling ahead. That tight feedback loop is why ReAct underpins so many tool-using AI agents, and it builds directly on chain-of-thought reasoning while adding the ability to act through tool calling.

Concrete example: ask a ReAct agent "Which of our two warehouses can ship a pallet of item SKU-22 today?" It thinks "I need stock levels for SKU-22," calls inventory(sku="SKU-22"), observes that warehouse A has 40 units and B has 0, thinks "only A qualifies," and answers from that real data — never inventing a number it did not look up.