Intermediate7 min read

What is ReAct?

ReAct ("Reason + Act") is the classic agent pattern where the model alternates between thinking out loud and taking actions with tools, using each result to inform its next thought.

Explain like I'm new to AI

ReAct stands for Reason + Act. It's a simple, powerful recipe for agents: instead of guessing an answer, the model repeats three moves:

Thought — reason about what to do next.
Action — use a tool (search, calculator, API).
Observation — read the tool's result.

…then thinks again, acts again, and so on until it can give a final answer. Step through a trace:

Question: "What is the population of the capital of France?"

Thought: I need the capital of France first, then its population.

ReAct interleaves reasoning (Thought) with tool use (Action) and results (Observation) — looping until it can answer.

Mental model

It's "show your work, then check it." Rather than blurting an answer, the model writes down its reasoning, does a concrete action to get real information, looks at what came back, and adjusts — like a detective narrating each deduction and following each clue.

How it works

The model produces a Thought (its reasoning).
It emits an Action — a tool call with arguments.
The system runs the tool and returns an Observation.
The observation is added to the context; the model thinks again.
Repeat until the model outputs a final answer instead of an action.

Interleaving reasoning with real observations keeps the model grounded — it reacts to actual results instead of hallucinating a path.

Real-world example

Question: "What's the population of the capital of France?" A pure guess might be wrong or outdated. With ReAct: Thought (need the capital) → Action search → Observation "Paris" → Thought (need its population) → Action search → Observation "~2.1M" → Answer. Each fact is looked up, not invented.

Technical explanation

ReAct (Yao et al., 2022) showed that interleaving chain-of-thought reasoning with acting beats either alone: reasoning helps plan and handle exceptions; acting grounds the reasoning in external facts, reducing hallucination. In practice it's implemented via function calling (the "Action" is a structured tool call) within the agent loop. Trade-offs: the reasoning traces cost tokens and latency, and models can still loop unproductively — so step limits and good tool design matter. ReAct is the foundation many richer patterns (planning, reflection, agentic RAG) build on.

Common mistakes

Common mistake

Letting the model "reason" its way to an answer it should have looked up. If a fact is external or time-sensitive, the Thought should lead to an Action (tool call), not a guess.

No loop limit, so Thought/Action cycles run on unproductively.
Poor tool descriptions, so the model picks the wrong Action.

When to use it

Tasks needing external information or multiple steps — the default, reliable agent pattern.

When NOT to use it

Simple questions the model can answer directly, where the extra reasoning/tool overhead isn't worth it.

Alternatives

Plan-and-execute (plan all steps first) and reflection add structure on top of ReAct's step-by-step style.

Quick quiz

Related concepts

The Agent Loop — The agent loop is the cycle of observe, reason, plan, and act that repeats — using tools and results — until the agent reaches its goal.
Planning & Reflection — Planning breaks a goal into sub-tasks before acting, and reflection lets the agent critique and fix its own work — together they make agents far more reliable.