Interview Questions

Agent engineering roles are among the fastest growing in technology. Interviewers know that demos are easy. So they test judgment. Here are the questions that come up, and how to answer them.

• Fast-growing roles
• Demos are easy
• They test judgment

How would you design a support agent?

Don't start with frameworks. Start with the risks. Ask what it may do, and what's the worst case. Say that you'd begin with a workflow, and add autonomy only where it's needed. Then cover tools, approval for risky actions, rules in code, evals from real tickets, and a staged rollout.

Your agent loops forever. What do you do?

Show a method. Read the trace. Is it repeating a tool? Often the tool's error is unclear, or its result doesn't help. Fix the tool first. Then add a limit on repeats, and a budget. Then add a case to your eval set, so that it can't come back.

How do you stop prompt injection?

Be honest. You can't fully stop it. So you limit the damage. Treat content as data. Remove one leg of the lethal trifecta. Give narrow tools. Gate actions in code. Get a human on risky actions. Red team it. If you claim that a better prompt solves it, you'll fail that question.

How do you measure an agent?

Offline, with a golden set, trajectory checks, and judges. Online, with task completion, escalation rate, cost per solved task, and human edit rate. Then explain how they connect. Offline evals block regressions, and online numbers show the true value.

Summary