Salesforce's Agent Fabric will pull failing AI agents offline when quality drops. Build that kill switch into smaller agents too

Salesforce just shipped the boring part of AI agents that most small builds skip: tests that run in production and a switch that turns the agent off.

What happened

Salesforce announced Agent Evals, API Evals and new cost controls for Agent Fabric, its control plane for managing agents across platforms. Salesforce says Agent Fabric handles over 1.2 billion requests a day.

• Agent Evals generates tests before deployment, then evaluates agents, LLMs and MCP servers in production against quality thresholds. When results fall short, it triggers predefined actions, including the Agent Kill Switch, with human approval where policy requires it.
• API Evals generates tests from API specifications to catch breaking changes before agents hit them.
• A FinOps Dashboard and Model Wallet set hard limits that stop requests before costs exceed budget. Salesforce says its cost policies cut token use by up to 75% in test scenarios.

Salesforce's own example: a retailer moves its order support agent to a cheaper model, scheduled tests find it giving wrong answers about delayed orders, and the agent is quarantined after approval until new instructions pass the tests again.

Agent Evals and API Evals enter beta in November 2026, with general availability in January 2027. The manual Agent Kill Switch and the cost controls are available now. Salesforce also says fewer than half of organizations with agents evaluate them on live traffic.

My take

You do not need Salesforce to copy this. The pattern is three parts: a fixed set of test cases with known answers, a schedule that runs them against the live agent, and an action when the score drops.

For a client support or lead qualification agent in n8n or Make, that can be 20 real past tickets with correct outcomes, run nightly. If accuracy falls below your line, flip a flag in a sheet or database that routes new requests to a human queue and pings the owner.

The model swap example is the one to remember. Most agent failures I see start with a change nobody tested: a cheaper model, a new prompt, an API field renamed. Run the tests after every change, not just at launch, and put a hard monthly cap on the API key itself.