3.4 — The ReAct loop (the engine of autonomy)

Once it holds its tools, how does the agent decide when to use them, without you telling it which button to click? Through the ReAct (Reason + Act) method — the base engine of any agentic loop (Yao et al., 2022). Unlike a script that follows a fixed path, the ReAct agent reasons at each step.

🟢 In plain words — ReAct is simply alternating “think” and “do”, in a loop. The agent tells itself what it’s going to do, does it, looks at the result, then starts over taking into account what it just learned — until it reaches its goal.

The loop:

1. 🧠 Reason — the agent analyzes its mission. “I need the weather in Paris. I’ll use my web search tool.”
2. ⚙️ Act — it uses the tool (it runs the “Paris weather” search).
3. 👀 Observe — it reads the result returned by the tool.
4. 🔁 Loop (adjust) — if it fails, it understands its error, corrects, retries. If it succeeds, it gives the final answer.

The agent repeats this loop until it succeeds.

This constant loop is what lets the agent handle the unexpected on its own.

🧩 Concrete example — mission: “book a free room tomorrow.” Reason: “I need to check the calendar.” Act: the agent opens the calendar. Observe: “2 PM is taken.” Loop: it retries with 3 PM, finds it free, books it, then confirms. No one told it “click on 3 PM”: it figured that out by looping.

⚠️ Common mistake — thinking a ReAct agent “knows” the answer in advance. It doesn’t: it moves forward by trial while reading the actual results. That’s exactly what lets it recover when a tool fails or returns the unexpected.

💡 2025-2026 note: ReAct remains the reference mental model. In practice, the frameworks (and recent reasoning models) automate much of this loop for you. Understanding ReAct means understanding what happens “under the hood” of any agent — including the no-code agents of lesson 3.8.

❓ Check your understanding — in what order do the 4 ReAct steps run?

See the answer

Reason → Act → Observe → Loop. Then it starts over until the goal is reached.

✅ In short

• ReAct = Reason → Act → Observe → Loop, repeated.
• The agent reasons at each step instead of following a fixed script.
• This loop lets it handle the unexpected and self-correct.
• It’s the shared engine of all agents, including no-code ones (3.8).