The ReAct framework

ReAct (Yao et al., 2022) interleaves Thought -> Action -> Observation steps: the agent produces a reasoning trace, takes an action (tool call), observes the result, and continues until it's ready to answer. Two advantages over pure CoT prompting: (1) actions let the agent pull fresh external information (search, APIs, code exec), grounding reasoning in real data; (2) each step's observation forces the next reasoning to be re-anchored to evidence rather than wandering. In production it has been superseded by structured tool-calling: a JSON-emitting model with typed tool schemas, which is more reliable than free-text action strings but trades off ReAct's inspectability. Senior nuance: function-calling ReAct also enables parallel tool calls, conditional tool selection, and post-tool reasoning verification. Many 2025-2026 agents use ReAct for the "think loop" but with tool-calling JSON instead of free-text actions.