Summary
ReAct (Yao et al. 2022): interleave thought, action, and observation in the agent loop. Most modern agents descend from ReAct. The pattern is simple and durable: think about what to do, do it, observe the result, think again.
Objectives
- 01State the ReAct loop.
- 02Explain what ReAct adds over plain tool calling.
- 03Recognize modern variants.
The Lesson
The loop
Thought: 'I need to find the current weather in Paris.' Action: search('weather Paris'). Observation: '15°C, cloudy.' Thought: 'I have the answer.' Action: finish('15°C, cloudy in Paris'). The explicit thought step is the key innovation.
Descendants
Reflexion adds critique loops on top of ReAct. Plan-and-execute separates planning from execution. Function-calling agents in modern APIs are ReAct with the thought/action structure implicit in the API shape. The core pattern is unchanged.
Key Ideas
- ReAct = thought → action → observation → repeat.
- Most modern agents are ReAct with better tooling.
References
- Yao et al., 'ReAct: Synergizing Reasoning and Acting in Language Models' (2022)