The Orchestrator

Not everything should be delegated. Before any build, the triage question is whether the work justifies the spend, and it has three tests.

Frequency times hours. A task done once a month that takes twenty minutes is not a delegation candidate at almost any token price; the build itself will cost more than a year of doing it by hand. A task done ten thousand times a week that takes ten minutes is worth automating even at frontier prices — the arithmetic shows up below. The triage is honest arithmetic, not enthusiasm: estimate the volume, estimate the current cost, divide.

The first-run trap. The first run of a new workload is always expensive — the preparation, the skills, the memory setup, the failed drafts — and counted honestly, the first change may cost about what it would have cost the ordinary way. What the investment buys is the capability to move fast on everything after. Chapter 4 made this point about time; it is equally true of money, and a workload assessment that only prices the steady state is lying by omission.

Experiment versus production. A $2 run is a $2 experiment. The same run ten thousand times a week is a budget line the finance department will eventually find. The discipline is to price the production version, not the demo: the real volume, the real failure rate, the real retry behavior, and the real model — not the one the demo happened to run on. A surprising share of “agent bills” that land in executives’ inboxes are experiments that were never re-priced when they became production.