Part 1 of 4 · 1 min read
Overview
Interviewers for agent roles use system design to see whether you connect model behaviour to infrastructure: tokens to cost, steps to latency, tools to security, runs to durability. The two prompts below are this level's interview questions: "Design a customer-support agent platform for many companies" and "How do you keep cost per resolved task under control?"
About the example, and what is real
Real: the capacity formulas (runs per day, peak rate, Little's law, workers, tokens per run with history re-send, cost, tokens per minute, bottleneck headroom), the design-review rubric, and the 17 test cases in code/mini-design. Assumed: every workload and price input (tenants, tickets, peak factor, steps, run times, tokens, $3 and $15 per million tokens, 10 percent cached-read price, a 2 million tokens-per-minute limit, 20 and 10 runs per worker, database limits). Not real: no system was built or load tested; the two designs are sketches reasoned from earlier modules.