ReAct — short for Reasoning + Acting — is the agent design pattern where the model interleaves explicit reasoning steps with tool-calling actions. Introduced in a 2022 Princeton/Google paper, ReAct asks the model to think out loud about what to do, take an action, observe the result, think again about what the result means, and repeat. The pattern is the conceptual ancestor of essentially every modern agent loop.
A basic ReAct trace alternates between three roles. Thought: "I need to find the user's recent orders." Action: search_orders(user_id=12345, days=30). Observation: "Found 3 orders, total $230." Thought: "The user asked about their last order specifically. Let me get its details." Action: get_order(order_id="ord_abc"). Observation: items, status: shipped, tracking number returned. Thought: "I have what I need to answer." Final Answer: "Your most recent order shipped on..."
Why ReAct works:
- The reasoning step constrains the action — by thinking out loud, the model is less likely to call the wrong tool with wrong arguments.
- The observation grounds the next step — the model sees what actually happened, not what it predicted would happen, before deciding the next move.
- The trace is debuggable — a human reviewing the trace can see exactly why the agent did what it did.
- It composes — long sequences of ReAct steps can solve multi-hop problems no single tool call could handle.
In 2026 the explicit ReAct format has been largely absorbed into LLM training. Modern frontier models (Claude Sonnet 4, GPT-5, Gemini 2.5) reason and act fluently inside the tool-calling APIs without needing the explicit "Thought / Action / Observation" scaffolding. But the conceptual pattern still drives how production agents are designed.
Variations and descendants:
- Plan-and-Execute — the model writes a multi-step plan upfront, then executes each step. Better for predictable workflows; worse when the right next step depends on prior observations.
- Reflexion — after a failed attempt, the model reflects on what went wrong and tries again with the lesson incorporated.
- Tree of Thoughts — branches over multiple reasoning paths, scores them and pursues the best.
- Voyager / Self-improving agents — the agent maintains and grows a library of learned skills across runs.
For a US team building agent products in 2026, the practical pattern is: rely on the model's native reasoning + tool-calling rather than scaffolding ReAct manually, but expose the reasoning trace in your UI for debuggability and trust. The "show your work" property of ReAct is what makes agents auditable, and that auditability is what makes them deployable in regulated industries.
