Anatomy of an Agent

Single LLM calls can't handle multi-step tasks. The loop enables sequential decision-making: each iteration sees the results of previous actions and adapts. Without a well-designed loop, agents either stop too early or run indefinitely.

The agent loop is like a GPS navigation system — it shows your current location (observe), plans the next turn (reason), you drive (act), then it recalculates based on where you actually ended up (update state).

Scenario

Task: 'Book a flight to Tokyo next Friday.' Step 1: search_flights('Tokyo', 'next Friday') → 3 options. Step 2: LLM picks cheapest → book_flight(id='JL408') → confirmation. Step 3: LLM returns booking confirmation to user. 3 iterations, done.

Solution

In Agent Foundations, apply Anatomy of an Agent to this scenario: Task: 'Book a flight to Tokyo next Friday. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Anatomy of an Agent (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Anatomy of an Agent happens in the code.

1def agent_loop(client, messages: list, tools: list, max_steps: int = 10) -> str:  # define a reusable function2    for step in range(max_steps):3        response = client.chat.completions.create(  # call the API4            model="gpt-4o", messages=messages, tools=tools,5        )6        msg = response.choices[0].message7        if not msg.tool_calls:8            return msg.content or ""  # return the result9        messages.append(msg)10        for tool_call in msg.tool_calls:11            fn_name = tool_call.function.name12            args = json.loads(tool_call.function.arguments)13            try:14                result = TOOL_REGISTRY[fn_name](**args)15            except Exception as e:16                result = f"Error: {e}"17            messages.append({18                "role": "tool",19                "tool_call_id": tool_call.id,20                "content": str(result),21            })22    return "Max iterations reached"  # return the result

• No max iteration limit
• Crashing on tool errors instead of feeding back to LLM
• Not streaming progress — user sees nothing for 30 seconds
• Sequential execution of independent tools
• No cost tracking — surprise API bills

Quick recap — the most important points from this module.