The news
Salesforce added cost controls and testing tools to Agent Fabric on 27 October 2026. Agent Fabric is its control panel for managing a company's AI agents, including agents that run on other platforms. The FinOps Dashboard shows AI spending in one view, Model Wallet sets hard limits that stop requests before a budget is exceeded, and Cost Optimization Policies reuse earlier answers, compress requests and trim replies. All three are available now. Salesforce says the policies cut token use by up to 75% in test scenarios.
Agent Evals writes tests before an agent goes live, then keeps checking agents, AI models and MCP servers in live use against quality levels the team sets. When an agent falls below them, a set action follows, such as taking it offline with the Agent Kill Switch, with a person's approval where policy requires. API Evals builds tests from API specifications to catch changes that break connections. Both are in beta from November 2026 and due for general release in January 2027. The Kill Switch itself is available now, and switching it on automatically from failed tests comes in January 2027.
The announcement gives no prices. Salesforce cites a LangChain survey that found fewer than half of organisations with agents test them on live traffic. The 75% saving comes from Salesforce's own test scenarios, not from customer bills.
What it means for you
- AI agents can give wrong answers and run up costs before anyone notices. The big vendors now build budget caps and quality checks into their tools.
- A hard limit stops requests once the budget is reached, so decide in advance what your customers see when that happens.
- Agent Fabric is built for large companies and has no published price, but the same habits work with any AI tool you pay for.
- The 75% saving comes from Salesforce's own tests, so measure it on your own traffic before you count on it.
What to do now
- Set a monthly spending limit for every AI tool or agent you pay for, with an alert before it is reached.
- Before moving an agent to a cheaper model, run it on real past customer questions and compare the answers.
- Keep a short list of questions with the right answers, and run it again after every change to an agent.
- Make sure someone can switch an agent off quickly, and that customers then reach a person instead.
How Didset can help
We build AI agents and CRM workflows with spending limits, tests and an off switch included from the first day.
Talk to us about it
