Technique
aka Reason+Act
ReAct
Agent pattern interleaving Thought, Action and Observation steps in a single trace.
Definition
ReAct (Yao et al. 2022) prompts the model to alternate reasoning steps ('Thought:') with tool invocations ('Action:') and observed results. The model decides when to act and when to think, producing more transparent and reliable agent traces than naive function calling.
Common use cases
- Research agents
- Coding agents
- Web-browsing bots