ReAct
Why ReAct Is Needed
Chain of Thought makes models "think before answering," but thinking alone isn't enough — many tasks also require "real action."
For example, "check the weather for me and tell me what to wear" — the model can't rely on imagination for real weather data. It needs to:
- Call search or an API to get real data
- Adjust judgments based on returned results
- Continue to the next step
CoT solves "how to think," ReAct solves "how to think while doing."
What Is ReAct
One-line definition: ReAct (Reasoning + Acting) is a working mode that makes models alternate between "think, act, observe" cycles.
Core loop:
Thought: What do I know now, what should I do next?
↓
Action: Call a tool or execute an action
↓
Observation: Look at what the returned result is
↓
Repeat: Continue to the next round until done or terminated
Analogy: Like doing research — you don't just sit and think through all the questions, but search, organize, and write as you go — whatever you find or discover affects your next decision.
Here, "Action" involves Tool Calling — letting the model issue calls that drive external systems to execute actions.
How to Do It: When to Use ReAct
Scenarios suited for ReAct:
- Needing external information: search materials, query databases, check real-time status
- Multi-step execution: read documents first, summarize, then write to a spreadsheet
- Changing environments: each step's result affects the next decision
- Long task orchestration: can't plan through to the end in one go, need to adjust as you execute
Scenarios not suited for ReAct:
- Single-turn text generation — adding this only increases complexity
ReAct vs CoT
| Dimension | CoT | ReAct |
|---|---|---|
| Core | Think only | Think + act alternately |
| Tool calls | No | Yes |
| Applicable scenarios | Math reasoning, logic reasoning | Complex tasks requiring research, action |
ReAct = CoT + Action + Observation
This explains why Agent frameworks universally adopt the ReAct pattern — Agents need not just "think clearly," but "think while doing while seeing results."
Remember this: CoT solves "how to think," ReAct solves "how to think while doing while seeing results."
Related terms: Chain of Thought · Tool Calling
Related Products