· Salesforce
Salesforce adds agent tests, hard spending limits and a kill switch to Agent Fabric
Salesforce announced on 27 October 2026 new testing and cost tools for Agent Fabric, its control plane for managing AI agents across platforms. Agent Evals checks agents against quality thresholds and can trigger an Agent Kill Switch, while new spending controls stop requests once a budget is reached.
What Salesforce announced
Agent Fabric is the Salesforce tool for finding, governing and monitoring AI agents, including agents built on other platforms. Salesforce says it handles more than 1.2 billion requests a day. The update adds three groups of features.
Agent Evals generates tests before an agent goes live and then checks agents, language models and MCP servers in production against quality thresholds the customer sets. When results fall short, it can start a predefined response, such as alerting an owner or pulling the agent offline through the Agent Kill Switch, with human approval where company policy asks for it. API Evals builds tests from API specifications so that a changed API is caught before it breaks an agent.
On cost, a FinOps Dashboard and a Model Wallet show AI spending in one view and set hard limits that stop requests before a budget is exceeded. Salesforce claims its Cost Optimization Policies cut token use by up to 75 percent, based on its own test scenarios. A Custom Agent Scanner extends automatic discovery to internal and third-party systems, and Open Policies apply one rule across API gateways from Azure, Apigee and Kong.
What it means for B2B sales teams
This is a tool for IT and AI leaders, not a sales app. It still matters to sales teams that now run AI agents for prospecting, inbound replies or meeting follow-up, often from two or three vendors at once. A prospecting agent that sends wrong facts to buyers, or burns through its budget in one week, is a sales problem first. Hard caps and an approval step before an agent is pulled offline are the kind of controls a revenue leader should ask for. In the EU, documented testing and human oversight also fit the direction of the AI Act. The limit: the main testing features are beta until January 2027, and the 75 percent saving is Salesforce's own test result, not a guarantee.
What to do now
- List every AI agent that talks to prospects or customers, which vendor runs it and who owns it.
- Set a monthly AI spending limit per agent and decide who can stop an agent when it misbehaves.
- Write five to ten test cases for each customer facing agent, such as pricing questions and objections, and rerun them after every model change.
- Keep a person on every reply that commits to price or scope; our sales strategy work starts with that review.
Sources
By Emre A. · Published