ReAct
An agent pattern that interleaves "reasoning" steps with "acting" steps — letting the model think out loud, take an action, observe, and reason again.
In plain English
ReAct (Reasoning + Acting) is the foundational agent design pattern: at each step, the model produces a "thought" (private reasoning), then an "action" (a tool call), then an "observation" (the tool's result). The loop continues until the model decides it has finished.
Why it works:
- Thought tokens give the model space to plan instead of jumping straight to an action
- Observations let the model react to real-world feedback (errors, unexpected results)
- Transparent traces — the chain of thoughts is human-readable and debuggable
Where it came from: Introduced in a 2022 Princeton/Google paper, ReAct was one of the first patterns to show that LLMs could reliably use tools. It's the conceptual ancestor of nearly every modern agent framework.
ReAct vs newer patterns:
- Plan-and-execute — plan first, then act (less adaptive but more efficient)
- Reflexion — add a self-critique step
- Tree of Thoughts — explore multiple reasoning branches in parallel
Most production agents are some variant of ReAct, often blended with planning or reflection.