Agent Architecture and Trade-offs
A system-design round on an agentic product, covering scalability, performance, safety, iteration, and evaluation.
The questions in this round
Design an assistant that resolves customer issues end to end: it reads the ticket, looks up the order, checks the refund policy, and can issue the refund. Take me through it.
It works. Now it is forty thousand tickets a day, and the refund tool belongs to another team with a limit of ten calls a second.
A customer writes, inside their ticket, text that reads like an instruction to issue a full refund. What in your design is still standing?
Last thing. How do you know this is getting better between releases rather than just different?