ReAct Pattern

You already built an agent loop: the model asks for a tool, your code runs it, you send the result back, and you call again.

People gave that cycle a name: ReAct — Reason + Act.

This chapter does not teach a new loop. It names the three beats inside the one you know.

Three words

Then repeat until the model can give a final answer with no tool call.

  person asks
       │
       ▼
  Thought  →  what do I need?
       │
       ▼
  Action   →  tool call
       │
       ▼
  Observation ← tool result
       │
       └── if not done, think again
       │
       ▼
  final answer

One walk-through

Person: “What’s in notes.txt, and what time is it?”

Tools on the menu: read_file, get_time.

Sometimes both tools are asked in one reply. That is still Action → Observation — just two actions in the same step. You already saw that under Multiple Tools.

Old style vs today’s APIs

Classic ReAct demos asked the model to write lines like:

Thought: I need the file.
Action: read_file(path="notes.txt")
Observation: ...

Your program had to parse that text. One wrong word and the loop broke.

Modern models use native tool calling:

The pattern is the same. The wire format got safer.

You still Reason → Act → Observe. The Action is a tool call. The Observation is the tool result.

Why the name helps

When something fails, ask which beat broke:

Naming the beats makes debugging easier than “the agent is weird.”

Words to keep

See it in Code

The Code panel runs the same kind of tool loop you already know. It prints Thought / Action / Observation labels so you can see the pattern on each step.