I rejected a candidate with “Built AI agents in production” on their resume because he couldn’t answer this Problem: Your agent can call tools that refund payments and send emails. How would you stop it from refunding something the user never asked for? Candidate response: I’ll just add a clear instruction in the system prompt that the agent should only call the refund tool if the user explicitly asks for a refund, and otherwise refuse it. But then I asked: The prompt can be overridden, What stops the actual refund call from going through? then silence..... What’s your response to get selected?