ReAct
ReAct is an agent pattern that interleaves reasoning with actions and observations. New tool results can change what the model does next.1
Tiny example
Section titled “Tiny example”decide: current weather is neededact: call the weather toolobserve: rain begins at 4 PMdecide: search for a morning trainact: call the train toolobserve: an 8 AM seat is availableanswer: recommend the morning optionThis differs from two fixed tool calls because the second action depends on the first observation.
FDE note
Section titled “FDE note”Use an allowlist, argument validation, a step limit, timeouts, and an escalation path. “Think, act, repeat forever” is not a production control loop.
Footnotes
Section titled “Footnotes”-
Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models”, ICLR 2023. ↩