Salesforce plans evaluation-driven shutdown controls for Agent Fabric

Salesforce announced a planned Agent Fabric workflow that would connect continuous evaluations to predefined interventions when an AI agent misses quality or policy thresholds. The change is meant to shorten the path from detecting a failure to alerting an owner, quarantining traffic or invoking a shutdown control.

The complete evaluation-to-intervention path is not generally available. Agent Evals and API Evals are scheduled to enter beta in November 2026 and reach general availability in January 2027. The Triggered-Agent Kill Switch is also planned for general availability in January. Salesforce did not give exact release dates.

The basic Agent Kill Switch, FinOps Dashboard, Model Wallet and Cost Optimization Policies are generally available now, according to Salesforce. That distinction means customers can already contain governed traffic and enforce spending limits, but cannot yet use the full production-evaluation workflow to trigger those interventions.

Agent Evals is designed to generate tests before deployment and evaluate agents, large language models and Model Context Protocol servers in production. Teams would define thresholds, and a failed result could initiate a preset response. The workflow could require human approval when an organization’s policy calls for it.

API Evals separately generates functional and performance tests from API specifications. Salesforce says those tests are intended to identify changes that could break an agent’s dependencies.

Salesforce illustrated the governance workflow with a hypothetical retailer whose order-support agent starts giving inaccurate information after a model change. In the scenario, Agent Evals flags the decline, requests approval to quarantine the agent, and lets the team revise instructions and rerun the evaluation before restoring service. Salesforce did not present the example as a production customer result.

The current Agent Kill Switch became generally available in September through Agent Fabric and Omni Gateway. Operators can contain governed traffic at the request, session, agent-instance or tenant level and later restore it. The mechanism blocks traffic through governed proxy paths; it does not establish that every process on an external hosting platform has been terminated.

The generally available cost layer includes a dashboard for spending visibility and Model Wallet limits that stop requests before an allocated budget is exceeded. Salesforce says its optimization policies can reduce token use by up to 75% in test scenarios through semantic caching, payload compression and response trimming. The company did not publish the test methodology, so the figure is a vendor claim rather than a guaranteed reduction in a customer’s total AI bill.

Custom Agent Scanner, which Salesforce says extends discovery to internal, behind-firewall and third-party environments, entered beta on October 27 and is planned for general availability in January 2027. Open Policies for Federated API Gateways, which applies policies across Azure, Apigee and Kong gateways, is generally available now.