Salesforce expands Agent Fabric with AI evaluations that will trigger a kill switch for failing agents

On October 27, 2026, Salesforce detailed the next phase of Agent Fabric, MuleSoft's platform for managing AI agents. The headline addition is a set of evaluations designed to spot underperforming agents and trigger corrective action. They join expanded discovery capabilities and spending and governance controls announced earlier. The VentureBeat article by Carl Franzen carefully separates what is new from what already existed.

Timeline. Agent Evals, which MuleSoft introduced at Dreamforce in September, will enter beta in November and become generally available in January 2027. API Evals, an upcoming tool for testing the software connections that agents depend on, follows the same schedule: beta in November and general availability in January 2027. A Custom Agent Scanner is also entering beta, with general availability set for January. The so-called Triggered-Agent Kill Switch is planned for the same month.

The underlying pitch is to shorten the gap between detecting a problem and fixing it. A company may know that an agent is giving wrong answers or consuming too many resources, yet still have to track down its owner, navigate another platform and get permission to suspend it. Andrew Comstock, senior vice president and general manager of MuleSoft at Salesforce, described customers' problems: “I have too many. The technologies are completely different. Nothing talks to anything. Everything works in silos.” Salesforce aims to manage those agents even when they were built on rival platforms: “It's not Salesforce-specific. It's for everything,” he said.

Agent Evals. At Dreamforce, the company had already described its ability to generate tests, measure results against success criteria and link failures to remediation actions. The October announcement adds dates and explains that the evaluations can go beyond pre-deployment testing to check agents, language models and Model Context Protocol servers in production. The goal is to detect behavior outside the organization's thresholds and initiate an appropriate response, with approval steps where policy requires them.

Salesforce's example is an illustrative retail scenario: an agent switches to a cheaper model and its answers about order status get worse. The evaluations detect the degradation and kick off a process to quarantine the agent, obtain approval if needed, and review and retest it. It is a hypothetical workflow, not a result from a real customer. Comstock added that evaluations can also help enforce expectations about tone: a customer service agent that is supposed to sound upbeat should not suddenly turn sarcastic. API Evals, for its part, will generate functional and performance tests from API specifications.

The kill switch. The basic feature is not new: MuleSoft introduced the concept in June and made it generally available in September. It uses gateway controls to block governed traffic, with the option to limit the intervention to a request, a session, an agent or a tenant, and to restore access afterward. The article highlights an important caveat: blocking an agent's access through a gateway does not necessarily mean terminating every process running on the external platform that hosts it.

What is being added to the January roadmap is the Triggered-Agent Kill Switch, which will link detected problems to a policy-driven intervention. For buyers, the practical difference lies between a control already available to halt governed activity and a future mechanism that ties evaluation results to that response.

Discovery and registry. Agent scanners were described in April as a way to identify agents in supported environments. The new beta piece is the Custom Agent Scanner, designed to extend discovery to internal systems, environments behind firewalls and additional platforms. Comstock said the registry does not require moving agents to Salesforce: “We don't move the agent; the agent keeps running there. We're just bringing in the pointer so you have a universal registry of all your agents.” An administrator connects the environments and brings their metadata into a shared registry. Nor is it necessary to use Agentforce, Salesforce's agent platform, to use Agent Fabric. Comstock also argued that the need depends more on the diversity of agent deployments than on company size.

Spending controls. The announcement highlights a FinOps Dashboard, Model Wallet budgets and policies to reduce model consumption, but not all of it is new: MuleSoft's August release already outlined model access, cost optimization, wallets and dashboard improvements. Model Wallet lets users set budgets in dollars or tokens per provider or model, with daily, weekly or monthly limits. Comstock stressed that visibility is not only about cutting spending, but about correlating it with the teams and projects that generate it.

The October release lists semantic caching, payload compression and response trimming among its optimization techniques, with token reductions of up to 75% in test scenarios. The article notes that this is a vendor test result, not a guarantee of a reduction in a customer's total AI bill.

Cross-vendor policies. Salesforce also highlights Open Policies, for applying governance across federated gateways such as Azure, Apigee and Kong. This is not entirely new either: MuleSoft described it in September, and the October release notes document the renaming of Universal Policies to Open Policies along with related improvements. Actual coverage depends on supported integrations and on traffic passing through enforceable controls. Comstock framed it as a prerequisite for broader deployment: “Today, governance is the real accelerator, because you're governed. You can answer questions better, feel safer and more comfortable, and that makes you go faster.”

Pricing. MuleSoft's public page lists a starting price of $2,000 per month for the overall portfolio, with Mule Credits used across capabilities. The agent management offering includes Agent Fabric features such as discovery, orchestration, governance and budgets, but directs buyers to sales. That figure should therefore not be presented as a confirmed price for the specific Agent Fabric configuration described. There are prepaid credit packages, a pre-commit option with an annual minimum of $50,000 and an unlimited option at custom pricing. No separate public price is set for Agent Evals or API Evals. The $2,000 budget Comstock used as an example in the interview was hypothetical.

Competition. Agent Fabric enters a growing market for organization-wide agent management. Microsoft Agent 365 is a close comparison in enterprise administration, with an agent registry, access and lifecycle controls, analytics and security integrations with Entra, Defender and Purview. ServiceNow's AI Control Tower likewise targets cross-vendor discovery, observability, governance, security and measurement. OpenAI Frontier offers a broader proposition that includes building and operating agents, connecting business systems, evaluating performance and managing permissions. Anthropic's Claude Managed Agents overlaps at the infrastructure layer but is a less direct comparison; its runtime costs $0.08 per active session hour, plus model and tool charges, and is not comparable to MuleSoft's pricing. The article clarifies that these are interpretations of published capabilities, not a claim that the products are interchangeable; a company might need both an agent runtime and a separate governance layer.

What can be used and what is missing. The most immediate item is the Custom Agent Scanner beta. The basic kill switch and the highlighted spending controls are already available. Agent Evals and API Evals arrive in beta in November and reach general availability in January 2027, along with the custom scanner and the Triggered-Agent Kill Switch. The central question for buyers is whether Salesforce can connect discovery, testing, spending visibility and enforceable intervention across the agents they actually use; upcoming releases will determine how much of that workflow functions as an integrated system.