Token volume stopped being a useful adoption metric when agents started running multi-step work for minutes or hours. More usage can mean more useful work. It can also mean retries, oversized context, the wrong model, or an agent continuing after the job has stopped creating value.
Recent product changes make the shift visible. AWS launched CloudWatch Coding Agent Insights with spend trends, token billing alerts, and cost-to-output comparisons. GitHub moved Copilot spend controls into its billing settings, including budgets, user-level controls, exports, and billing API access. Cost governance is becoming part of the agent operating layer.
Tokens are a meter, not an outcome
The risk is already concrete. Financial Times reporting, summarized by Tom's Hardware, described an internal Amazon project that used Claude Sonnet to match author details with product listings. The reported bill reached $1.8 million—860% above its allocated budget—and the issue was detected roughly five months after it began.
That is a reported internal example, not evidence that every agent deployment will overrun. It still exposes a control failure that can recur anywhere: spend was visible too late and was not bounded tightly enough to the value of the task.
A token count can explain part of a bill. It cannot show whether the workflow produced a useful pull request, resolved a customer case, reconciled a finance exception, or removed a manual handoff. Adoption and value are different measures.
Put the budget on the workflow
Agent budgets should not live only in a central infrastructure dashboard. Each deployed workflow needs a named business owner and an operating envelope.
The owner should know which job the agent is allowed to perform, which models it can call, what a successful output looks like, and when the run must pause. Engineering or platform teams can maintain telemetry and routing. Finance can set commercial boundaries. The workflow owner remains accountable for whether the spend produces a result worth keeping.
This is also why a flat organization-wide token cap is too blunt. A high-value research workflow and a repetitive formatting task should not inherit the same model, context window, or escalation path. Route capability and budget according to the job.
Build a minimum cost-control loop
Before expanding an agent beyond a pilot, define:
- Owner: one person accountable for the workflow's value and exceptions.
- Budget: a per-run, daily, or monthly limit that matches the way the work operates.
- Alert: an early threshold that reaches the owner before the hard cap is exhausted.
- Output signal: a verifiable unit such as an accepted code change, reviewed report, resolved case, or completed reconciliation.
- Routing rule: the smallest model and context that can meet the required quality, with escalation only when justified.
- Pause rule: a clear stop condition for repeated failures, unusual token growth, or cost without accepted output.
- Review cadence: a recurring comparison of spend, quality, cycle time, and human rework.
The goal is not to minimize every token. It is to stop unowned consumption. A useful agent can be expensive and still earn its place. What it cannot be is expensive without an owner, a boundary, or evidence of output.
Related services: Finance, Engineering, Implementation