ReAct has the model think, act, then read results before thinking again.
DefinitionWhat it means
ReAct (Reason plus Act) is a prompting pattern in which the model alternates between a Thought step that plans what to do next, an Action step that invokes a tool, and an Observation step that reads the tool's output, repeating this loop until it has enough information to answer.
Why it mattersWhy you should care
ReAct is the backbone of most agent loops shipping today because it makes multi-step tool use auditable: each thought and observation is logged, so builders can debug why an agent took a wrong turn and product teams can show users a trace of how an answer was reached.
At a glanceSee it
What actually ends the loop — a sufficiency check after each observation plus a step budget that caps runaway cycles, with the two classic failure modes hanging off the exits.
Why ReAct exists — it interleaves the reasoning trace of chain of thought with the tool calls of an act-only agent, curing the blind spots of each.
Where you see itIn the wild
- Agent framework loops in LangChain, LlamaIndex, and custom orchestration code.
- Debug traces and logs when diagnosing why an agent looped or stalled.
- System-design whiteboards for agent architecture.