ReAct

A reasoning pattern where an agent alternates explicit thought ("Reason") and action ("Act") steps until the task completes.

ReAct (Reasoning + Acting) was introduced by Yao et al. in 2022 and has become the dominant agent loop pattern in production deployments. The model emits an explicit thought (“I need to look up the customer’s recent orders”), then an action (a tool call), reads the result, emits a new thought, and continues until done.

The pattern’s strength is debuggability: the trace shows the agent’s reasoning step-by-step, which makes failure modes inspectable. Its weakness is verbosity — each cycle burns tokens, and aggressive ReAct loops can spiral into expensive over-thinking on tasks that should resolve in 2-3 actions.

Most production agent harnesses ship ReAct as the default loop pattern. Frameworks like LangChain and CrewAI implement ReAct under the hood; you don’t usually configure it directly. When you tune agent cost, you’re often tuning ReAct loop length implicitly.