- Salesforce announced Agent Evals, API Evals and cost controls for Agent Fabric. Several pieces are available now, while the evaluation-triggered kill switch is due in January 2027.
- Salesforce says fewer than half of organizations with agents test them on live traffic. Stopping a failing agent often means finding its owner first.
- Before buying, ask which controls ship today, which are betas, and whether blocking an agent at a gateway actually stops it.
A failing agent is a people problem before it is a technical one
An AI agent that gives wrong answers rarely announces itself. Salesforce says as much in its latest announcement for Agent Fabric, its control plane for managing agents across platforms. Agents "can fail silently," it says, giving wrong answers and running up costs before anyone notices.
Salesforce points to two gaps. The first is testing. By its account, under half of the organizations that run agents check them against live traffic. The second is action. Even when a team finds a fault, it must first locate whoever owns the agent and win permission to limit it. Meanwhile the agent keeps working and keeps spending.
Here is our view. Detection is only half the problem. The other half is the delay between finding a fault and being allowed to act. The announcement stresses that second half. Fewer than half of organizations test agents on live traffic, and Salesforce says stopping a faulty agent still means tracking down its owner. Software that narrows the second gap swaps a chase for approvals with a rule written in advance.
What Salesforce announced
Salesforce led with three groups of features. The first is Agent Evals. It writes tests automatically before an agent goes live. Once the agent is running, it scores the agent against quality thresholds. It does the same for the language models and the Model Context Protocol servers the agent uses. That protocol is the standard that lets agents connect to outside tools.
A poor score can set off a response the team defined in advance. One such response is the Agent Kill Switch. Where company policy demands it, a person approves first.
The second is API Evals. It builds tests from API specifications. Salesforce's logic is simple. When an API that an agent relies on changes, the agent can break.
The third is a set of cost controls. A FinOps Dashboard and a Model Wallet sit at the centre. Salesforce describes the wallet as hard limits that stop requests before spending passes budget. Cost Optimization Policies cut token use by up to 75%, Salesforce says, "based on test scenarios." That figure comes from tests, not from customers in production.
Read the dates before the headline
VentureBeat's coverage separates what is available from what is promised. Agent Evals and API Evals both enter beta in November 2026. Both reach general availability in January 2027. The Custom Agent Scanner is in beta now and is due for general availability in January 2027.
The kill switch needs the closest reading. Salesforce lists the Agent Kill Switch as generally available today. VentureBeat reports the basic version became available in September. The version triggered by a failed evaluation, called the Triggered-Agent Kill Switch, is due in January 2027.
So the automatic link between "quality dropped" and "agent stopped" is the part buyers cannot use yet. VentureBeat also notes that the September Dreamforce announcement already described connecting failed evaluations to corrective actions. Much of this release adds dates and extensions to earlier announcements.
How the kill switch works, and what it does not do
VentureBeat reports that the switch uses gateway controls to block governed traffic. An intervention can be scoped to a request, a session, an agent or a tenant. Access can be restored afterward.
That design has a limit, which VentureBeat spells out. Blocking an agent's access through a gateway does not necessarily end every process on the outside platform that hosts it. A blocked agent can be cut off from your systems and still be running somewhere else.
Salesforce's retail example shows the intended flow. A retailer moves its order-support agent to a cheaper model. Scheduled tests find wrong answers about delayed orders. The system asks for approval to quarantine the agent. The team fixes its instructions and retests before going back live. VentureBeat stresses this is a hypothetical workflow, not a reported customer result.
Why a vendor-neutral control plane matters
Salesforce says 96% of enterprises run agents on more than one platform. Andrew Comstock leads MuleSoft and Agent Fabric at Salesforce. He told VentureBeat that customers say: "Nothing talks to each other. Everything runs in a silo." He added that the approach is "not Salesforce specific."
One customer, GE Appliances, described the appeal as a single place to see its agents, who built them, and how to shut them down. That points to a lesson. As agents multiply, the weak point is as much ownership and authority as the models. In the scenario, a cheaper model gives worse answers, and the retailer chooses to have a person approve the quarantine.
What to ask before you buy
First, ask which controls are generally available today and which are betas. Get the January 2027 dates in writing.
Second, ask what a kill switch actually stops. If it blocks traffic at a gateway, find out whether the agent keeps running on its host platform.
Third, name an owner for every agent now, whatever the tooling. The announcement's own diagnosis is that incidents drag on while someone looks for the person who can say yes.
Fourth, decide where a human must approve a shutdown and where a rule may act alone. Salesforce says teams can require approval where policy demands it. That choice belongs to your organization, not the vendor.
An off switch is only useful if someone is allowed to press it, and if it reaches the thing that needs stopping.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Salesforce.
