The controls are arriving in stages
Agent Evals, first presented at Dreamforce in September, is meant to test agents against defined success criteria and connect failures to corrective action. Salesforce says the checks can run before deployment or on working agents, language models and Model Context Protocol servers. If an agent crosses an organization’s thresholds, the system can trigger actions that include approval steps.
Salesforce’s retail example is hypothetical, not a report from a customer. An agent moved to a cheaper model begins giving worse answers about order status. An evaluation detects the decline, the agent is quarantined, and, if required, someone approves the next step before the agent is fixed and tested again.
The planned API Evals tool addresses a different target: it will generate functional and performance tests from API specifications. Both products are scheduled to enter beta in November, with full launches planned for January 2027.
The kill switch itself is not new. MuleSoft introduced the concept in June and made the basic capability generally available in September. It uses gateway controls to block managed traffic, with intervention scoped to a request, session, agent or client; access can later be restored.
Blocking an agent’s gateway access does not necessarily stop every process running on the external platform where that agent operates. The future Triggered-Agent Kill Switch is meant to connect detected problems with rule-based intervention. Salesforce discussed linking failures to corrective actions in September; the October announcement gives that mechanism a planned January release.
Finding agents without moving them
Custom Agent Scanner is entering beta to extend discovery into internal systems, environments behind firewalls and additional platforms. MuleSoft had already described scanners for supported environments in April; general availability for the custom scanner is planned for January.
Customers will not have to move agents onto Salesforce. Andrew Comstock, MuleSoft’s senior vice president and general manager at Salesforce, told VentureBeat that agents can keep running where they are while Agent Fabric collects links to them in a shared registry. Administrators connect the environments and add their metadata. Agentforce, Salesforce’s own agent platform, is not required.
Comstock said the coordination problem depends more on how varied an organization’s deployments are than on its size. A business using several frameworks and vendors can struggle to manage its agents without being a large enterprise.
That makes Agent Fabric’s central promise less about running agents than about governing a mixed fleet. Salesforce says customers can manage agents built on competing platforms, not just its own products.
Cost controls are part of the same pitch
Salesforce also points to FinOps Dashboard, Model Wallet budgets and rules intended to reduce model consumption. Those capabilities were introduced across earlier announcements: in September, MuleSoft described setting dollar or token budgets for a provider or model, with daily, weekly or monthly limits.
The October release description lists semantic caching, data compression and shorter responses. Salesforce says these methods reduced token use by up to 75% in test scenarios. That is a vendor test result, not a promise of an equivalent reduction in a customer’s overall AI bill.
I think the more consequential test is whether these pieces work as one operating layer. Discovering an agent, seeing its spend, judging its behavior and stopping its managed traffic are separate capabilities; their value depends on how reliably Agent Fabric connects them across the environments customers actually use.
The announcement is less clear about what that control will cost. MuleSoft’s public pricing page lists a starting price of $2,000 per month for a broader product portfolio, but directs customers to sales for agent-management offerings. That figure is not a confirmed standalone price for the Agent Fabric configuration announced in October. The page also lists prepaid Mule Credits, a $50,000 minimum annual commitment option and custom pricing for unlimited use. Salesforce does not publish a separate price for Agent Evals or API Evals, and the available information is not enough to estimate the cost of a particular evaluation workload.
The market is splitting between control and execution
Agent Fabric enters a market where products overlap, but do not necessarily do the same job.
Anthropic lists Managed Agents Runtime at $0.08 per active session hour, excluding models and tools. That is not directly comparable with MuleSoft’s portfolio starting price or a custom Agent Fabric contract. The products are not interchangeable: depending on existing infrastructure and integration support, a company may need both an environment to run agents and a separate layer to govern them.
My guess is that Salesforce’s strongest case will be with companies that already have agents spread across platforms and need a shared view more than another place to build them. But the announcement leaves the operational boundary unresolved: how much of an agent’s activity can the gateway actually control when the agent keeps running elsewhere?
The basic kill switch and the named spending controls are available now. Custom Agent Scanner is the nearest new beta; Agent Evals and API Evals are scheduled for beta in November, with general availability planned for January 2027. The scheduled releases will show whether Salesforce can turn a collection of controls into a dependable layer across agents it does not host.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X