Salesforce said on October 27 that its Agent Fabric control plane now includes an Agent Kill Switch, a Model Wallet with hard spending limits and a FinOps Dashboard, all generally available today. The evaluation features that decide when to pull an agent offline, Agent Evals and API Evals, enter beta in November 2026 and reach general availability in January 2027.
What Salesforce announced
According to Salesforce’s announcement, Agent Fabric is its control plane for managing and governing agents across platforms, and it handles more than 1.2 billion requests a day. The company says the new pieces are meant to catch quality failures and runaway token spend before customers notice them.
Salesforce lists five additions. Agent Evals generates tests before deployment, then evaluates agents, LLMs and MCP servers in production against quality thresholds. When results fall short, predefined interventions fire, including the Agent Kill Switch, with human approval where a team’s policy requires it. API Evals builds functional and performance tests from API specifications to catch breaking changes. The FinOps Dashboard and Model Wallet give budget owners one view of AI spending and hard limits that stop requests before costs exceed budget. Cost Optimization Policies use semantic caching, payload compression and response trimming, which Salesforce says cut token usage by up to 75% in test scenarios. Two discovery features, a Custom Agent Scanner and Open Policies for Federated API Gateways, extend coverage to internal environments and to gateways from Azure, Apigee and Kong.
The release schedule matters more than the feature list
Salesforce’s availability section splits the launch in two. Generally available today: the Agent Kill Switch, the FinOps Dashboard, Model Wallet, Cost Optimization Policies and Open Policies for Federated API Gateways. Beta today: the Custom Agent Scanner, with general availability in January 2027. Beta in November 2026 and generally available in January 2027: Agent Evals and API Evals. The Triggered-Agent Kill Switch also arrives in January 2027.
That sequencing means a customer can use the Kill Switch and cap spend this week. The automated loop Salesforce describes, where a failing evaluation quarantines an agent without a person in the middle, will not exist as a generally available product until the new year. Salesforce’s own retailer scenario shows the intended flow: a model swap lowers cost, scheduled tests detect wrong answers about delayed orders, and the system asks for approval to quarantine the agent. That scenario depends on the January release.
The numbers Salesforce cites
Two statistics in the post frame the problem. Salesforce says fewer than half of organizations with agents evaluate them on live traffic, and that 96% of enterprises run agents on more than one platform. The post gives no survey name, sample size or date for either figure, so treat both as Salesforce’s characterization of the market. The 75% token reduction is also Salesforce’s figure, from its own test scenarios, and the post does not describe the workload.
The multi-platform figure explains the product’s positioning. Salesforce says Agent Evals works “no matter what platform the agent is built on,” which makes Agent Fabric a layer above any single vendor’s agent runtime. A buyer whose agents run on Salesforce, a hyperscaler and a third-party service gets one place to set a threshold, if the integrations behave as described.
How the spend control compares with Google and AWS
Salesforce is the third vendor in three weeks to describe a platform-enforced limit on agent spending, and each puts the limit at a different level. On October 8, Google described a Gemini agent that stops when a project reaches an administrator’s cap, which we covered in Google’s Gemini agent with real-time spend caps and agent identities. AWS then described a per-session budget set outside the model, in its agent payments account. A project cap bounds the total, and a session cap bounds what one runaway task can spend.
Salesforce’s Model Wallet gives budget owners hard limits, but the post does not say whether a limit applies per project, per agent or per session. It also does not say who can raise a limit mid-task, or what an agent does when a request is refused. The Google and AWS caps cover agents running on those vendors’ own platforms. Salesforce’s pitch is a limit that spans platforms.
What it means for the buying committee
Three questions belong in the next vendor review.
- Which agents are in scope today? The Custom Agent Scanner is in beta, so a company with agents behind a firewall or on a third-party platform should test discovery before counting on a complete registry.
- Where does the limit sit? Ask Salesforce whether Model Wallet limits attach to a budget owner, an agent or a session, and compare the answer with the project and session caps already on the table from Google and AWS.
- Who approves a quarantine? Salesforce says approval is configurable. Decide before January whether a failing evaluation may pull a customer-facing agent offline on its own, and who is on call when it does.
Companies already running agent estates should also read the announcement next to Dataiku’s product for counting the agents a company is running, since discovery and cost control depend on the same inventory.
What to do this quarter
CIOs and Chief AI Officers can use the gap between today’s general availability and January’s automation as a planning window. Turn on the FinOps Dashboard now to learn what agents actually cost, set Model Wallet limits conservatively for the first agents and write the quarantine policy before the November beta so the evaluation thresholds have an owner when they arrive. Customer quotes in the post from Capita, GE Appliances and Newell Brands describe Agent Fabric as a central place to see and govern agents. None of the three reports a savings figure or an incident avoided, so a business case needs your own baseline.
Source: Salesforce Newsroom
