Salesforce is detailing the next phase of its Agent Fabric platform, with evaluations designed to identify malfunctioning AI agents and initiate corrective action, alongside expanded discovery capabilities and previously announced spending and governance controls.
The October 27 announcement sets a November beta and January 2027 general-availability target for Agent Evals, which MuleSoft first announced at its annual Dreamforce conference in September. It also outlines API Evals, a forthcoming tool for testing the software connections agents depend on, starts a beta for custom agent scanners and schedules an evaluation-triggered agent kill switch for January.
For enterprise buyers, the announcement combines forthcoming capabilities with extensions and existing tools. The basic kill switch became generally available in September, while MuleSoft outlined budget controls, spending dashboards and token-saving features over the summer. The September announcement also described connecting failed evaluations to corrective actions.
The broader proposition is to shorten the distance between discovering a problem and addressing it. A company might know that an agent produces incorrect answers or consumes too many resources, but still need to identify its owner, navigate another platform and obtain permission to suspend it.
In an interview with VentureBeat, Andrew Comstock, senior vice president and general manager of MuleSoft at Salesforce, described familiar problems behind the company’s approach.
“I have too many of them. The technologies are totally different. Nothing talks to each other. Everything runs in a silo,” Comstock said, describing customer challenges that informed Agent Fabric.
Salesforce wants to manage those agents even when customers build them on competing platforms.
“I’ll emphasize, it’s not Salesforce specific. It’s across everything,” he said.
Agent Evals moves toward beta
Salesforce had already described Agent Evals at Dreamforce in September, including generating tests, measuring outcomes against success criteria and connecting failures to remediation.
The October announcement adds rollout dates and describes how evaluations can extend beyond testing before deployment to checking agents, language models and Model Context Protocol servers in production. The goal is to identify behavior that falls outside an organization’s thresholds and initiate an appropriate response, including approval steps where policy requires them.
In Salesforce’s illustrative retail scenario, moving an agent to a cheaper model produces worse answers to order-status questions. Evaluations detect the decline and initiate a process to quarantine the agent, obtain approval where required, and revise and retest it. That is a hypothetical workflow, rather than a reported customer result.
Comstock said evaluations can also help companies enforce expectations about how agents communicate. A customer-facing agent that should sound upbeat, for example, should not unexpectedly become sarcastic.
The forthcoming API Evals tool addresses another part of the system: generating functional and performance tests from API specifications. Both evaluation products are scheduled to enter beta in November and become generally available in January 2027.
The kill switch already exists; evaluation-triggered intervention is coming
The kill switch itself is not new to this announcement.
MuleSoft introduced the concept in June and made the basic capability generally available in September. It uses gateway controls to block governed traffic, with options to scope an intervention to a request, session, agent or tenant and restore access afterward.
That distinction matters: blocking an agent’s access through a gateway does not necessarily terminate every process running in the external platform that hosts it.
What Salesforce now puts on the January roadmap is a Triggered-Agent Kill Switch, connecting detected problems to policy-driven intervention. The September evaluation preview had already discussed linking failures to remediation; the October announcement supplies a more specific release target for the triggered capability.
For buyers, the practical distinction is between an available control for stopping governed activity and a forthcoming mechanism that connects evaluation results to that response.
Existing discovery gets a custom extension
Agent discovery also predates today’s announcement. MuleSoft described its agent scanners in April as a way to identify agents across supported environments.
The addition entering beta today is a Custom Agent Scanner, intended to extend discovery to internal systems, environments behind firewalls and additional platforms. General availability is planned for January.
Comstock said the registry does not require customers to relocate their agents to Salesforce.
“I’ll emphasize, we’re not moving the agent; the agent’s still running there,” he said. “We’re just pulling the pointer so that you can get a universal registry across all of your agents.”
An administrator connects the relevant environments and brings their metadata into a shared registry. Customers do not have to use Agentforce, Salesforce’s own agent platform, to use Agent Fabric, Comstock said.
He also argued that the need depends more on the diversity of a company’s agent deployments than its size. A business using several agent frameworks and providers may encounter coordination problems even without being a large enterprise.
Spending controls build on the summer roadmap
The announcement also highlights a FinOps Dashboard, Model Wallet budgets and policies intended to reduce model consumption.
These are not all newly introduced features. MuleSoft’s August release overview outlined model access, cost optimization, wallets and dashboard improvements. A September explanation of Model Wallet described setting dollar or token budgets by provider or model, with daily, weekly or monthly limits.
The controls address a familiar tension: an individual developer may choose an expensive model to get a task done faster, while the organization needs to understand the combined bill.
Comstock said visibility should extend beyond reducing total spending to identifying the teams and projects responsible for it.
“So it’s not just about keeping the cost down. That’s obviously part of it, but it’s also making sure that you can correlate that with the right teams and projects who are actually driving changes in your organization,” he said.
Salesforce’s October release describes semantic caching, payload compression and response trimming among its optimization techniques. It claims token reductions of up to 75% in test scenarios. That is a vendor test result, not a guaranteed reduction in a customer’s overall AI bill.
Cross-vendor gateway policies also predate today
Salesforce is also emphasizing Open Policies for applying governance across federated gateways, including Azure, Apigee and Kong.
The underlying capability is not making its first appearance today. MuleSoft described Open Policies in September, while its October release notes document the renaming of Universal Policies to Open Policies and related enhancements.
For enterprises, the attraction is applying common controls across infrastructure they already operate. Actual coverage still depends on the supported integrations and where traffic passes through enforceable controls.
Comstock framed that governance as a prerequisite for broader deployment.
“Nowadays, governance is your actual accelerant because you’re governed. You can answer better questions. You can feel more confident and comfortable, and that makes you go faster,” he said.
MuleSoft lists a $2,000 starting price, but evaluation pricing remains unclear
Salesforce’s public MuleSoft pricing page lists a starting price of $2,000 per month for the broader portfolio, with Mule Credits used across capabilities.
The page’s agent-management offering includes Agent Fabric functions such as discovery, orchestration, governance and budgets, but directs buyers to contact sales. The $2,000 figure should therefore not be presented as a confirmed standalone price for the specific Agent Fabric configuration described in this announcement.
The pricing material outlines prepaid credit packages, a precommit option with a $50,000 annual minimum and a custom-priced unlimited option. It does not establish a separate public price for Agent Evals or API Evals, or enough detail to calculate what a particular evaluation workload would cost.
The $2,000 project budget Comstock used as an example during the interview was hypothetical, not a product-price disclosure.
Microsoft, ServiceNow and OpenAI compete for enterprise agent oversight
Agent Fabric sits in a growing market for managing agents across an organization, although competing products combine oversight and execution in different ways.
Microsoft Agent 365 is a close comparison on enterprise administration. Microsoft describes an agent registry, access and lifecycle controls, analytics and security integrations with products such as Entra, Defender and Purview.
ServiceNow’s AI Control Tower similarly targets discovery, observation, governance, security and measurement across vendors. Its announced capabilities include controls for risky activity and visibility into AI costs and business outcomes.
OpenAI Frontier is another relevant competitor, with a broader proposition that includes building and operating agents, connecting business systems and context, evaluating performance, and managing permissions. The overlap is in enterprise agent management, even though the scope is not identical to Agent Fabric’s emphasis on governing agents across existing environments.
Anthropic’s Claude Managed Agents overlaps at the infrastructure layer but is a less direct comparison. It provides managed facilities for running agents, including execution environments and orchestration. Agent Fabric’s central pitch is discovering and governing agents that may continue running elsewhere.
Anthropic publishes a Managed Agents runtime price of $0.08 per active session-hour, in addition to applicable model and tool charges. That is not directly comparable with MuleSoft’s portfolio starting price or a negotiated Agent Fabric contract.
These distinctions are an interpretation of the vendors’ published capabilities, rather than a claim that the products are interchangeable. An enterprise could need both an agent runtime and a separate governance layer, depending on integration support and its existing infrastructure.
What buyers can use—and what remains ahead
The most immediate addition in this announcement is the Custom Agent Scanner beta. The basic kill switch and highlighted spending controls are already generally available.
Agent Evals and API Evals are scheduled for beta in November, with general availability in January 2027. The custom scanner is also scheduled for general availability in January, alongside the Triggered-Agent Kill Switch.
For buyers evaluating the announcement, the central question is whether Salesforce can connect discovery, testing, spending visibility and enforceable intervention across the agents they actually use. The forthcoming releases will determine how much of that workflow can operate as a connected system.