Core agents
What is ReAct?
ReAct is a pattern that interleaves explicit actions with observations.
The original paper describes reasoning-and-acting loops in language models. In products, the observable parts are the actions and observations.
Visual
ReAct loop
Interleaving reasoning (thought) with action (tool call) and observation.
Why it matters
ReAct (Reasoning + Acting) formalized the pattern that most production agents use: interleave a reasoning step with an action step. The model thinks about what to do, does it, reads the observation, and repeats.
The original academic paper named the pattern, but the idea predates it. What matters is the separation between reasoning (inspectable thoughts) and acting (tool calls with observable results).
ReAct in practice
In production, ReAct is usually implemented as a tool-calling loop where the model emits both a thought and a tool call on each iteration. The thought explains the intent; the tool call executes it; the observation grounds the next thought.
Frameworks like LangChain and LlamaIndex implement ReAct as a default agent pattern.
Key takeaways
- 1ReAct separates reasoning from action — both are inspectable.
- 2The pattern is the default in most agent frameworks.
- 3Observations ground the model's next decision in evidence.
Common mistakes
- ✕Using ReAct for tasks that don't require tools — a simple prompt suffices.
- ✕Not logging the thought steps, making the trace incomplete.
Related terms
Concept neighborhood
Terms linked from ReAct in the glossary graph.