Part 1 of 6 · 1 min read
Overview
Level 7 made one agent's context cheaper to manage. Level 8 asks the opposite question: when is it right to have more than one agent at all? The answer is a trade of latency and context size against tokens and information loss, and the trade depends on the shape of the task. An interviewer will want a decision rule, not a pattern catalogue.
About the example, and what is real
Everything runs offline in code/mini-fanout. Real: the dependency graphs, the longest-path computation (levels), the token and latency accounting for the single agent, for each worker and for the orchestrator, the prompt-cache price model from module 7, and the 14 tests. Assumed: every size (system prompt 1,500 tokens, shared background 3,000, 700 tokens per step, 6 steps per sub-task), the 25 percent extra steps workers take because they lack the whole picture, the 3 percent chance of losing a fact per handoff, and the hazard rates for reliability (error per step grows with tools and with context). The direction of the hazard rates comes from module 6; the values are mine. The reliability column is a parametric model, not a measurement: it shows the shape of the trade-off, and the "robust model" row shows how the answer reverses when the rates change. No model is called.