CSE Roadmap

Common tier · about 5 hours · 13 parts

Part 9 of 13 · 1 min read

Answering trade-off questions in an interview

Why did you choose a single agent over a multi-agent design?

"On the evidence I know of, a multi-agent team scored lower on simple reasoning (9.3 against 9.7) and used 43 times the tokens, but higher on tool-heavy tasks (9.2 against 6.8). So it depends on traffic mix: it wins above about 14 percent tool-heavy. For my workload I measured X percent, so I chose Y, and I would switch if that share passes Z." Follow-up: "How reliable is that comparison?" One author, one model, ten tasks, an LLM judge: a prompt for my own measurement, not a law.

How do you make a decision when the criteria conflict?

"I apply hard requirements first, drop dominated options, then compare the rest with a break-even or a weighted score on a fixed scale, and check how sensitive the winner is to the weights and the measurements." Follow-up: "What if two options tie?" Then the weights are not the issue; I measure the quantity I assumed.

How do you present a trade-off to a non-technical stakeholder?

"As a price: 'the faster design costs about two points of quality on hard cases; it is worth it if speed matters more than that.' I show one chart with the break-even and say which numbers are measured." Follow-up: "What if they disagree with your weights?" Show the critical weight; if their weights are inside the range where the decision holds, it does not matter.

What would make you reverse a decision?

"A measurable trigger written in the decision record: for example, if measured planner quality is more than 0.04 below the large model, or if p95 stays above 8 seconds after the change." Follow-up: "How often do you revisit?" When a model, price, traffic mix or requirement changes, and on a schedule.