1. The agent design interview

An agent design interview is a round where you get an open task, such as "design an agent that handles refund requests for our online store", and have to design an LLM agent for it and defend each choice. You already know the basic loop: the model picks a tool, the tool's result goes back into the conversation, and the loop repeats until the task is done (LLM Engineering 101: Agents covers it). The round tests the judgment around that loop: what the agent may do, how much freedom it gets, and what stops it from doing damage.

The basic principle is to work in order. Pin down the task, especially which actions cannot be undone, decide how much autonomy it earns, then make the design decisions in a fixed sequence. Use this framework for "design an agent that does X", for "walk me through an agent you built", and for "would you build an agent at all?" Skip it for a question about one mechanism, such as how ReAct (the reason-then-act loop) works. For a fixed sequence of known steps, plain code with one or two LLM calls beats an agent.

Mid-level candidates are expected to clarify before drawing, pick a sensible autonomy level, and list the tools. Senior candidates attach controls to the specific actions that cannot be undone, name a success measure they could actually compute, and argue against extra autonomy by naming a concrete way it would fail. We use one running example throughout, a customer-support agent for an online store that answers order questions and can issue refunds and reship orders, and we deliberately stop before it becomes a full architecture. The chapters follow the order you would use in the room.