ReAct Prompting

ReAct Prompting is a framework that structures LLM outputs as alternating Thought, Action, and Observation steps.

Each cycle lets the model reason about its next move, call a tool or API, then update its plan based on what comes back. This loop is the conceptual foundation behind most AI agent architectures. Also known as: ReAct Pattern, Reason and Act

What this topic covers

Foundations — ReAct Prompting structures LLM outputs as a repeating Thought–Action–Observation loop, giving the model a way to reason, act on external tools, and correct course rather than producing a single static response.
Implementation — The guides here cover implementing a ReAct loop from scratch, choosing between a hand-rolled pattern and native function calling, and the trade-offs that decide which approach fits a given use case.
What's changing — Native tool calling has absorbed many ReAct use cases, but understanding the pattern remains essential for debugging agent behavior and evaluating whether newer abstractions actually solve the underlying problem.
Risks & limits — ReAct-based agents can take irreversible external actions without explicit user consent at each step—understanding where accountability sits and how to constrain action scope is essential before deploying these systems.

This topic is curated by our AI council — see how it works.