Salesforce Unveils Agent Evals, Cost Caps for Agents

Salesforce has added hard spending limits and token-saving policies to Agent Fabric, its control plane for managing AI agents across platforms. It also announced Agent Evals, which tests agents and can pull one offline when its quality slips, but that will not reach beta until November.

The Salesforce announcement on 27 October covers testing for agents and for the APIs they rely on, cost controls, and wider discovery of agents running elsewhere. The cost features are generally available now. The evaluation tools follow in beta in November 2026, with general availability in January 2027.

In brief

Agent Evals generates tests, checks agents against quality thresholds and can trigger the Agent Kill Switch. Beta in November 2026; generally available in January 2027.
API Evals builds tests from API specifications to catch breaking changes, on the same timetable.
The FinOps Dashboard, Model Wallet spending caps and Cost Optimization Policies are generally available today.
Salesforce says the policies cut token usage by up to 75% in test scenarios.

What Agent Evals and API Evals do

According to Salesforce, Agent Evals will “automatically generate tests before deployment, then evaluate agents, LLMs, and MCP servers in production against defined quality thresholds”. When results fall short, they “trigger predefined interventions”, including the Agent Kill Switch, “with approval where the team’s policy requires it”. MCP servers are the connectors agents use to reach tools and data. Salesforce’s Agent Fabric product page currently describes MCP server and LLM evaluations as “coming soon”.

API Evals deals with a quieter failure. “Agents fail when the APIs they depend on change,” Salesforce says. The tool “automatically generates functional and performance tests from API specifications, catching breaking changes before they reach production”.

Salesforce illustrates this with a retailer that switches its order-support agent to a cheaper model. In that example, scheduled tests catch wrong answers about delayed orders and the agent is quarantined after a human approves.

“Enterprises can only scale agents as fast as they can keep them accurate and on budget,” said Andrew Comstock, SVP and general manager of MuleSoft and Agent Fabric.

Cost controls and what is available now

The cost tools are the part customers can use today. Salesforce says 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 “cut token usage by up to 75% (based on test scenarios)” through semantic caching, payload compression and response trimming. Salesforce has not published the test details behind that figure.

Salesforce’s availability list splits as follows:

Generally available today: Agent Kill Switch; Open Policies for Federated API Gateways, which apply one policy across Azure, Apigee and Kong gateways; and the FinOps Dashboard, Model Wallet and Cost Optimization Policies.
Beta today, generally available January 2027: Custom Agent Scanner, for finding agents in internal systems and third-party platforms.
Beta in November 2026, generally available January 2027: Agent Evals and API Evals.
Generally available January 2027: Triggered-Agent Kill Switch.

Salesforce cites “over 1.2 billion requests handled daily” on Agent Fabric but does not say what counts as a request. Its announcement adds that it “may include references to services or features that are still in development and are unreleased”, and that customers “should make their purchase decision based on fully released and available features”.

Why Salesforce is linking tests to the off switch

“What excites us most about Agent Evals is that it doesn’t stop at telling us something is wrong,” said Poonam Deshpande of ASICS Digital, a beta partner. Salesforce’s pitch is that the same system should act on the result: a failed evaluation triggers a pre-agreed response, from alerting an owner to quarantining the agent. It cites a LangChain survey finding that “fewer than half of organizations with agents evaluate them on live traffic”.

The timing matters. The kill switch is generally available now, but the Triggered-Agent Kill Switch and general release of Agent Evals, whose failed tests are meant to set off interventions, arrive only in January 2027. Until then, the automatic loop Salesforce describes is something customers can try in beta, not run in full production.

Rivals are moving in the same direction at different speeds. IBM this week made agent identity generally available but kept its own agent kill switch in tech preview. ServiceNow’s AI Passport, which limits what agents and MCP servers may do at runtime, was announced for the fourth quarter of 2026 and was not yet available when TSN reported it. Salesforce pitches Agent Fabric as working across platforms, including ones it does not own, and its new scanner and gateway policies extend that reach.

The cost controls are available now, so customers can check them directly. The 75% figure is harder to judge: it is an upper bound from Salesforce’s own test scenarios, and real savings are likely to depend on how much of a company’s agent traffic can be cached or trimmed. Salesforce, which recently agreed to buy Braze, is selling the control plane as the way to keep agents both accurate and affordable.

What happens next

Agent Evals and API Evals enter beta in November 2026. They, the Custom Agent Scanner and the Triggered-Agent Kill Switch are due to become generally available in January 2027. Salesforce’s own caveat applies: unreleased features may change, and it advises customers to buy on what is already available.