Part 1 of 4 · 1 min read
Overview
Agent cost is multiplicative (steps x context x model price) and has a fat tail (loops). The project criterion for this level is dashboards that show cost per run. Interviewers ask "how do you keep cost per resolved task under control?" and expect budgets, caching, routing with verification and a quality floor, not just "use a cheaper model".
About the example, and what is real
Real: the cost arithmetic, history-resend cost model, caching model, router and cascade logic, budget class, cache simulations and the 19 test cases in code/mini-cost. Assumed (all invented, not price lists): $0.25 / $1.25 and $3 / $15 per million input / output tokens for "small" and "large", cached input at a tenth of the price, the success curves (small: 0.97 minus 1.05 x difficulty; large: 0.99 minus 0.25 x difficulty), task difficulty distribution, classifier accuracy, verifier rates, loop rate, similarity distributions, query repeat distribution. Not real: no API calls, no real tokenisation, no real quality measurement.