Part 1 of 4 · 1 min read
Overview
An agent design question is a budget question. Each extra model call adds seconds, dollars and a chance of failure, and a loop multiplies them by a random number of rounds. Interviewers watch whether you can reason about that budget with numbers and decide what to cut.
About the example, and what is real
Real: the composition engine (sequence, parallel, loop, cache, route), retry and timeout handling, Monte Carlo percentiles, sensitivity analysis and 21 tests in code/mini-sysdesign. Assumed: every component latency, cost, availability and quality number; the 20% cache hit rate; the 60% router share; the loop distribution (1, 2 or 3 rounds with probabilities 0.5, 0.3, 0.2); independent failures; log-normal latency shapes fitted from a median and a 95th percentile. Not real: no real system or vendor was measured, and the design is fictional.