Agent Fabric’s spend limits ship now. Auto shutdown waits

The claim

On 27 October Salesforce announced that it “Expands Agent Fabric to Automatically Stop Runaway Costs and AI Agent Failures Before They Impact Customers” (vendor, Salesforce). Agent Fabric is the MuleSoft control layer that registers, governs, and meters agents across platforms, including ones that do not run on Salesforce.

The test is simple: which parts of that sentence can a buyer switch on this quarter?

The inputs

Every status below is as Salesforce or MuleSoft states it (vendor), with the date each feature was first announced.

The result

The “automatically” in the headline depends on Agent Evals deciding that an agent has failed and the Triggered-Agent Kill Switch acting on it. Both are 66 to 96 days away if the January date holds (our maths). The kill switch you can use today is the manual one, which had already been generally available for 55 days on announcement day (our maths).

Assumptions: “January 2027” means any day in that month, and the beta and GA dates hold. Salesforce’s own post says customers should buy on released features only.

The cost half stands. Budget owners can set hard limits that block requests once a budget would be exceeded, and see AI spend in one dashboard, today (vendor).

Where it breaks

• The gateway is the boundary. The kill switch blocks traffic across model proxies and to other agents (vendor, MuleSoft). A gateway block does not necessarily stop every process running on the external platform hosting the agent (independent, VentureBeat).
• Coverage follows the traffic. Open Policies only govern where supported integrations exist and traffic passes through an enforceable control point (independent).
• The 75% is a test result. It comes from Salesforce test scenarios, not a customer bill (vendor).
• The retailer story is illustrative. The example of evals catching a cheaper model’s wrong answers is a scenario, not a reported customer result (independent).
• Rivals already ship overlapping tools. Microsoft Agent 365, ServiceNow AI Control Tower, and OpenAI Frontier cover registry, governance, or evaluation (independent). A firm whose agents mostly live on one of those may not need Salesforce as the control plane.

For Salesforce and MuleSoft owners this quarter

1. Step 1. Set Model Wallet budgets per agent, not per org, so one runaway agent hits its own ceiling first.
2. Step 2. Rehearse the kill switch. Stop and restore one non-production agent, and confirm the audit record lands where your security team reads it.
3. Step 3. Keep the platform-side stop. For agents hosted outside Salesforce, the runbook still revokes the credential at the source.
4. Step 4. Join the November Agent Evals beta with your own thresholds. Decide what “failed” means for each agent before the vendor default decides it for you.
5. Step 5. Get the price in writing before any plan assumes the January features.

Owner: the Salesforce platform owner with FinOps. If you want a second pair of hands on the runbook, that is the work our Salesforce practice does.

Missing

• A public price for Agent Fabric, Agent Evals, or API Evals. VentureBeat reports MuleSoft’s portfolio starts at $2,000 a month, with a $50,000 annual precommit option (independent); neither is a confirmed Agent Fabric price.
• A named customer result for evaluation-triggered shutdown.
• Which external agent platforms the kill switch fully stops, rather than cuts off at the gateway.