Salesforce connected evaluation to shutdown: a failing agent no longer keeps serving while someone looks for it
Direct answer
On October 27, 2026, Salesforce detailed the next phase of Agent Fabric, its control plane for managing agents across platforms. The central change is that evaluation stops being only a report: Agent Evals generates tests before deployment and evaluates agents, models and MCP servers in production against thresholds defined by the customer. When a result falls below the threshold, the system starts a predefined intervention, including the agent kill switch, with an approval step when team policy requires it.
What arrives and when
- Agent Evals: beta in November 2026 and general availability in January 2027. It was first previewed at Dreamforce in September and now adds dates and production evaluation.
- API Evals: generates functional and performance tests from the specifications of the APIs agents depend on, to catch changes that would break them before production.
- Triggered-Agent Kill Switch: scheduled for January 2027. It connects an evaluation result to a policy-driven intervention, such as quarantining an agent. The basic kill switch has been generally available since September.
- Custom Agent Scanner (beta): extends discovery to internal systems, environments behind firewalls and additional platforms.
- Spending controls: budget policies that stop requests before they exceed the limit. The company says its controls cut token use by up to 75% in test scenarios. That figure comes from company test scenarios, not a guaranteed saving.
Silent failure, with an example labeled as hypothetical
- The scale of the problem: per the announcement, Agent Fabric handles more than 1.2 billion requests a day and 96% of enterprises run agents on more than one platform. At that volume, a silently failing agent accumulates wrong answers and cost while the team locates the owner and obtains permission to suspend it.
- Salesforce's example: a retailer switches its order-support agent to a cheaper model and scheduled tests detect that the agent starts giving inaccurate information about delayed orders. The evaluation starts the process to quarantine it, obtain the required approval and revise and retest it. Salesforce presents that flow as an illustrative scenario, not as a customer-reported result.
- The lesson for your business: before scaling agents, define three things in writing: the quality threshold that triggers action, who approves the intervention, and what happens to traffic while the agent is quarantined. Without those three answers, evaluation is a dashboard reviewed after the damage.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- What to do when your AI agent makes a mistake
- How to tell if your AI is working: the 5 KPIs that matter
- What an AI agent is and how a real business uses one
Frequently asked questions
What is Salesforce Agent Evals?
An Agent Fabric tool that generates tests before deployment and evaluates agents, models and MCP servers in production against thresholds defined by the customer.
When will it be available?
In beta in November 2026 and generally available in January 2027, together with the evaluation-triggered kill switch.
Does the kill switch already exist?
The basic kill switch has been generally available since September. What is new in January 2027 is connecting it to evaluation results and intervention policies.
What does API Evals do?
It generates functional and performance tests from API specifications, to catch changes that break agents before production.
