Salesforce has added hard spending limits, an agent kill switch and automated quality tests to Agent Fabric, the control plane it sells for managing AI agents across platforms.
The October 27 announcement covers three groups of features. Agent Evals generates tests before deployment and then evaluates agents, LLMs and MCP servers in production against quality thresholds. API Evals builds functional and performance tests from API specifications to catch breaking changes. The cost controls are a FinOps Dashboard, a Model Wallet that sets hard limits stopping requests before spend exceeds budget, and Cost Optimization Policies that Salesforce says cut token usage by up to 75% in test scenarios.
The Agent Kill Switch, the FinOps Dashboard, Model Wallet and Cost Optimization Policies are generally available now. Agent Evals and API Evals enter beta in November 2026 and reach general availability in January 2027.
Why it matters
Salesforce cites two figures for the problem. Fewer than half of organizations running agents evaluate them on live traffic, and 96% of enterprises run agents on more than one platform. The example in the announcement is a retailer’s order-support agent, and the release does not mention marketing agents.
Marketing teams that run agents to launch campaigns, as described in Marketing Campaigns Are Starting to Run Themselves, face the same two exposures: an agent that is wrong without anyone noticing, and a bill that grows between invoices. We reported on HubSpot’s default spend limits in Ask Every AI Vendor for a Default Spend Cap in Writing, and Salesforce’s Model Wallet is the same idea applied at the gateway level, across vendors.
One original angle
A limit enforced at a gateway applies only to traffic that passes through the gateway. In our reading, a marketing operations lead should ask which of the team’s agents route through Agent Fabric, and which run inside a vendor’s own product with its own meter. Salesforce’s own 96% figure says most enterprises run agents in more than one place. Agent Evals is still in beta, so for now the hard limits are the part a buyer can switch on.
Source: Salesforce
