A single API bill for fifty agents is not a cost model. Five dimensions turn aggregate spend into something you can act on — select each to see the question it answers.
One number, one account, aggregated across every team, agent, and tenant. Nobody can say which agents are responsible. It is a bill, not a cost model — and you cannot manage what you cannot attribute.
The baseline dimension: which agents cost what, and how that has changed over time.
This is the question that should drive decommissioning decisions and almost never does — because the cost side of the ROI calculation is simply missing. Agents accumulate, and without per-agent spend nobody can argue for retiring one.
Within a single agent, task types have very different economics. An agent handling both simple status lookups and complex multi-system reconciliations has a bimodal cost distribution, and the average conceals both modes.
Task-level attribution is what reveals that 6% of requests drive 60% of spend — which is exactly where routing and caching investment belongs.
Required for chargeback, and required much earlier than teams expect once an agent serves multiple internal customers.
Retrofitting tenant attribution is painful, because it means threading identity through code that was never designed to carry it. Add it before you need it.
Which models are actually being used, versus which your architecture document claims. Routing logic drifts and fallbacks fire more often than anyone realises.
A fallback to a frontier model triggering on 30% of requests instead of the intended 2% is a common and expensive silent failure — invisible without this dimension.
The dimension almost nobody instruments, and the most valuable one. Cost of successful completions versus cost of failures, retries, and abandoned tasks.
Spend on failure is pure waste, and it is frequently 15–30% of the total. It is also the number that makes the business case for reliability work — because it converts an engineering argument into a financial one.
Contains runaway loops. Belongs in the agent framework, not in monitoring — by the time an alert fires, the money is spent. Set it from your observed distribution, around the 99th percentile of normal executions, so it catches pathology without killing legitimate hard cases.
Protects the shared budget from one misbehaving agent, and gives you a blast-radius boundary aligned with your security boundaries.
Prevents one consumer exhausting shared capacity — the same reasoning as API rate limiting, applied to cost rather than throughput.
Fall back to a cheaper model and accept lower quality. For high-volume, low-stakes work.
Defer until the budget window resets. For asynchronous processing.
Stop and hand off to a human. For anything consequential.
An agent that hits a ceiling and returns a degraded answer without signalling has converted a cost control into a correctness problem. This is a genuine autonomy border, so it belongs in the agent's design record rather than in infrastructure config nobody reads.
Includes the cost of the failures. This is the figure to compare against the manual process it replaces, and it is usually meaningfully worse than the per-call number quoted in business cases. It also improves when reliability improves, which aligns the incentive correctly.
What a sponsor can actually evaluate — and the only view that reveals whether this gets cheaper per unit as volume grows. Agents with heavy per-task reasoning frequently do not benefit from volume the way traditional automation does, and knowing that before committing to a rollout is worth a great deal.