The Structure of a Turn

Four steps, every time

Every turn does the same four things: call the model with the messages so far, inspect why it stopped, act by running any tools it asked for, and append the assistant message and the tool results. Then the loop starts again. If you can name these four steps, you can place any guard exactly where it belongs.

One turn of the loop

Each turn calls the model, checks why it stopped, runs any requested tools and appends the results.

The skeleton

call_model, run_tools and final_text are placeholders for your own code. Notice the loop has no way to end except a clean stop or the for limit.

def run(messages, max_steps=10):
    for step in range(max_steps):
        reply = call_model(messages)               # 1. call
        messages.append({"role": "assistant", "content": reply.content})
        if reply.stop_reason != "tool_use":        # 2. inspect
            return final_text(reply)
        results = run_tools(reply)                 # 3. act
        messages.append({"role": "user", "content": results})  # 4. append
    raise RuntimeError("step limit reached")

Put guards between steps

Checks for budget, time and repetition belong right after "inspect" and before "act". That way a runaway loop is stopped before it spends money on another tool call.

Quick check: Which step of a turn runs the tools the model asked for?

• Call
• Act
• Append
• Inspect