Salesforce, Snowflake, and Atlassian Shipped the Agent Budget Cap. Kill It as the Control.

Three control planes shipped the same control in one week, and none of them is the control. This is a Kill List entry for the agent budget cap, with the line that should replace it.

The short version

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, Agent Evals that quarantine an agent below a quality threshold, and a kill switch, over 1.2 billion requests a day. Snowflake's Cost Governance, per Nick El-Rayess, attributes AI spend to a team or an agent and enforces hard limits in near real time, with weekly per-user quotas and anomaly alerts across up to 50 spend cohorts. Atlassian's Ryan Narragon cites Gartner's 150,000 agents per Fortune 500 enterprise by 2028 and the 13% who believe their governance is right. Every one of these fires at the wallet, after the spend, and Salesforce's own worked example shows where the bill was actually 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. Kill the hard cap as the control. Replace it with a named list of cost-bearing keys per agent, a re-quote routed to the budget owner whenever a change sets one, written in cost per completed task. Keep the wallet as the backstop.

Three vendors, one week, one control

Salesforce went first, on October 27. The Agent Fabric expansion adds Agent Evals, which generate tests before deployment and then evaluate agents, models, and MCP servers in production against quality thresholds, triggering interventions up to the Agent Kill Switch with approval where policy requires it; API Evals, which generate tests from API specs so an agent does not fail when a dependency changes; and the cost side: a FinOps Dashboard and Model Wallet that give budget owners one view and hard limits that stop requests before costs exceed budget, and Cost Optimization Policies that reduce token usage by up to 75% in test scenarios through semantic caching, payload compression, and response trimming. The post cites a LangChain figure that fewer than half of organizations evaluate agents on live traffic, and says Agent Fabric handles over 1.2 billion requests a day. Agent Evals is beta in November and general availability in January; the wallet and the policies are available now. Andrew Comstock's line is that enterprises can only scale agents as fast as they keep them accurate and within budget.

Snowflake went last, on November 3, in Nick El-Rayess's Expedition post. Cost Governance, generally available, attributes AI spend to a team, department, agent, or application and enforces hard limits in near real time to stop runaway agents before they exceed budgets. Weekly per-user quotas. Org-level budgets in dollars or credits. Tag-based anomaly detection with daily alerts and up to 50 custom spend cohorts per account. Dynamic Model Routing, in private preview, with up to 3x greater token efficiency than a frontier-only path in internal testing. And, in private preview, Agent Trace Intelligence, which uses traces to find agents wasting money in loops, tools that repeatedly fail, and the real cost of each completed task.

Atlassian sat between them, on October 30. Ryan Narragon's post argues that agent sprawl is an identity and accountability risk rather than a shadow IT problem, because unapproved agents use the organization's real data and permissions. He cites Gartner's figure that an average global Fortune 500 enterprise will have more than 150,000 agents in use by 2028, and that only 13% of organizations believe they have the right agent governance in place, and recommends tiering actions by risk, keeping permission inheritance as the baseline, and tracking agent counts and invocation patterns so teams can consolidate and narrow scope.

The same day as Salesforce, Vercel's changelog said agent sessions now cost 67% less. I read the headline only, and it is here because it is the other half of the week: the price per session fell by two thirds while three vendors shipped tools to stop the bill.

The bill is set before the wallet sees it

All three catch the money at the wallet. A hard limit stops the request when the budget is gone. An anomaly alert fires when the daily number is off. A quota ends the user's week. Each of these is downstream of the spend. What none of them watches is the change that set the rate.

Salesforce's own worked example is the clearest statement of the problem, and it is in the announcement. A retailer moved its order-support agent to a cheaper model. Scheduled tests after the model change found the agent giving inaccurate information about delayed orders. Agent Evals flagged it and requested approval to quarantine the agent, the team revised the instructions, re-ran the evaluation on the lower-cost model, and returned the agent to production. The post presents this as cost savings pursued safely, and it is. Read it for the order of events. The model swap was the cost decision. It was made by a person, in a change, before any eval ran. The thing the control plane caught was the quality failure, afterwards, on a schedule.

Now run the same story with the arrow pointing the other way. Someone moves the agent to a better model, or raises the retry count, or widens the context window, or adds a second tool call per step. Nothing fails. The quality tests pass, because the agent got better. The wallet sees the bill twelve days later and stops requests at the budget line, good runs and bad alike, and the alert says the money is gone. It does not say which line moved.

I had this from the other side

The Scope Guard Caught 6 of 9. Name the Keys That Carry the Bill. is the same shape at the scale of one engineer. Lesson 05 of Becoming Forward Deployed Engineer ships a scope policy as a file, paths you may edit, no deletes, no new dependencies, a cap on files per change, forbidden strings, and a guard that checks each proposed change against it. On ten labeled changes it catches six of six mechanical violations and zero of three changes that break no rule and would still end the engagement.

Two of those three misses are a person's job, as the lesson says. The third is not. It is top-k raised from 5 to 20 inside a file the policy allowed. That is a cost change, and cost-bearing keys are a finite list the course's own lesson 04 had already named. So the fix was a seventh field, cost_bearing_keys, and a rule: any added line that sets one of them is reported as a re-quote and routed to the name on the cost line of the handshake. Six of nine became seven of nine.

That is the control the three vendors did not ship. It sits at the change, not the wallet. It fires before the first run at the new rate, not after the budget. And it names the line.

Kill the cap as the control

The Kill List entry is the hard budget cap as the thing a team points at when asked how agent cost is governed.

A cap fires at the end of the money, days after one line moved the price per task. It stops the good runs with the bad, because a wallet cannot tell a loop from a busy Tuesday. And its alert carries the least useful fact available, that the budget is gone, instead of the most useful one, which change set the rate. It is a smoke alarm sold as a thermostat.

The replacement has three parts.

A named list of cost-bearing keys per agent. Model, top-k, retries, context size, tool fan-out, schedule. The list is short, finite, and different per agent, and writing it down is the whole of the work. Kill the Per-Call Agent Meter made the same point about pricing: the meter on the call hides the keys that move the call count.

A re-quote on any change that sets one. Routed to the budget owner before the change ships, the way a scope guard routes a forbidden path. Not an approval gate on every change; a re-quote on the ones that touch the list. Atlassian's risk tiers are the right shape for this, with cost-bearing keys as a tier of their own.

Written in cost per completed task. This is the unit Snowflake's Agent Trace Intelligence says it will report, and it is the right one, because it is what the buyer pays against. Tokens are the vendor's unit. 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. Only the per-task number tells the owner which one just happened.

The wallet stays. A hard cap is the right backstop for the failure nobody predicted: a loop, a runaway schedule, a tool that started charging. Snowflake's loop detection and Salesforce's kill switch are that backstop. The kill is calling it the control.

Under Enterprise AI Agents the claim is that what matters is how many deployed agents still complete production work after ninety days, and what each successful outcome cost. A cap protects the first number by cutting the agent off. A re-quote on the line that moved protects the second, which is the one the renewal is priced on.

This week, for each agent in production: write its cost-bearing keys as a list, open the last thirty days of merged changes, and mark every one that set a key on that list. The ones with no re-quote attached are the bill you have not seen yet.

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

Sources: Salesforce, "Salesforce Expands Agent Fabric to Automatically Stop Runaway Costs and AI Agent Failures Before They Impact Customers," Salesforce Newsroom, October 27, 2026, read in full; the LangChain figure is as Salesforce cites it. Nick El-Rayess, "Governed Context, Security and Spend Controls for AI Agents," Snowflake blog, November 3, 2026, read in full. Ryan Narragon, "Tackling agent sprawl across the enterprise," Inside Atlassian, October 30, 2026, read in full; the Gartner figures are as he cites them. Vercel, "Vercel Agent sessions now cost 67% less," changelog, October 27, 2026, headline only. Falk Gottlob, "The Scope Guard Caught 6 of 9. Name the Keys That Carry the Bill." (October 14, 2026) and "Kill the Per-Call Agent Meter."