ReAct is a prompting pattern where the model alternates between reasoning and actions, using observations from tools to drive the next step.
Core Loop
- Thought: decide what to do next.
- Action: call a tool (search, code, DB, etc.).
- Observation: read the tool result.
- Repeat until a stopping condition.
Why It Matters for Agents
ReAct is the default mental model for tool-using agents: planning is not done once upfront; it is continuously updated based on real-world feedback.
Notes
- Many frameworks implement a ReAct-like loop even when they do not explicitly label it as such.
- In 2025+ models increasingly internalize ReAct behavior, but explicit prompting still matters for reliability.
Further Reading
- ReAct: Synergizing Reasoning and Acting in Language Models --- Yao et al. (2022)