What is wrong with the cap as the control?

It fires at the end of the money, days after one change moved the price per task, and every run in between was billed at the new rate. It stops the good runs with the bad, because a wallet cannot tell a loop from a busy day. And its alert carries the least useful fact, that the budget is gone, rather than which line moved. Salesforce's retailer example has the order of events: the model swap was the cost decision, the quality failure was found later by a scheduled test.

What are cost-bearing keys?

The short, finite, per-agent list of settings that move cost per task: model, top-k, retries, context size, tool fan-out, and schedule. In The Scope Guard Caught 6 of 9, adding that list as a seventh policy field and reporting any change that sets one as a re-quote routed to the name on the cost line took Lesson 05's guard from six of nine catches to seven of nine, because the miss that cost money was top-k raised from 5 to 20 inside a file the policy allowed.

Why cost per completed task and not tokens?

Because the buyer pays per task and the vendor bills per token. A model swap that halves tokens and doubles retries is flat per task; a top-k change that leaves tokens per call alone and quadruples calls per task is a four-times re-quote. Snowflake's Agent Trace Intelligence, in private preview, says it will report the real cost of each completed task, which is the right unit for the re-quote.

The wallet, as the backstop for the failure nobody listed: a loop, a runaway schedule, a tool that started charging. Snowflake's loop detection and Salesforce's kill switch are exactly that. The kill is naming the cap as the control when asked how agent cost is governed.

Is a Hard Budget Cap the Right Cost Control for AI Agents?

Is a hard budget cap the right cost control for AI agents?

THE SHORT ANSWER

As the backstop, yes. As the control, no. In one week Salesforce's Agent Fabric added a Model Wallet with hard limits that stop requests before costs exceed budget, cost policies that cut tokens by up to 75% in tests, and Agent Evals that quarantine an agent below a quality threshold; Snowflake's Cost Governance, per Nick El-Rayess, enforces hard limits in near real time with weekly per-user quotas and anomaly alerts across up to 50 spend cohorts; and Atlassian's Ryan Narragon cited Gartner's 150,000 agents per Fortune 500 enterprise by 2028 against 13% who believe their governance is right. Every one fires at the wallet, after the spend, and Salesforce's own worked example shows where the bill was set: a retailer swapped its order-support agent to a cheaper model and a scheduled test found the wrong answers afterwards. My version from the other side is Lesson 05's scope guard, six of nine, where the miss that cost money was top-k raised from 5 to 20 inside an allowed file, and a seventh field listing the cost-bearing keys made it seven of nine. The control is a named list of cost-bearing keys per agent (model, top-k, retries, context size, tool fan-out, schedule), a re-quote routed to the budget owner whenever a change sets one, written in cost per completed task rather than tokens. The cap stays as the backstop for the loop nobody predicted.

As the backstop, yes. As the control, no. The control is the re-quote on the change that sets a cost-bearing key, written in cost per completed task, with the wallet behind it.

The three announcements, the retailer example, and the key list are in Salesforce, Snowflake, and Atlassian Shipped the Agent Budget Cap. Kill It as the Control.. The six-of-nine guard and the seventh field are in The Scope Guard Caught 6 of 9. Name the Keys That Carry the Bill., and the pricing version of the same argument is Kill the Per-Call Agent Meter.

Last reviewed 2026-11-03 · 1 min read

Is a hard budget cap the right cost control for AI agents?

SOURCES

THE LONG VERSION

RELATED ANSWERS