<- All papers

ReAct: Synergizing Reasoning and Acting in Language Models

Combines reasoning traces with environment actions so agents can plan, observe, and revise while solving tasks.

ReAct is the agent workflow reference for the site. It makes a simple point durable: useful agents need an action loop, but they also need observable state and repair paths.

Builder Notes

  • Model the loop as plan, act, observe, reflect, and verify.
  • Keep actions inspectable so failures can be traced.
  • Pair ReAct-style workflows with evaluation logs, not only prompt examples.