.png)
Salesforce Agent Fabric: What the New Controls Do and What Your Team Still Owns
Salesforce just gave Agent Fabric a kill switch, hard spending limits, and a way to test agents in production. Some of it works today. Some of it arrives in 2027. None of it works without setup your team has to do first.
On October 27, 2026, Salesforce announced Agent Evals, API Evals, and new cost controls for Agent Fabric, its AI control plane for managing agents across platforms. The Agent Kill Switch, Model Wallet, FinOps Dashboard, and Cost Optimization Policies are generally available now. Agent Evals and API Evals enter beta in November 2026 and reach general availability in January 2027.
Here is the catch. MuleSoft's setup guide says some of these controls need an Omni Gateway, an identity provider that issues JSON Web Tokens, and registered agents.
Quick answer: Agent Fabric can help monitor agent behaviour, control AI spending and contain problematic agents. Getting these controls working requires the right setup: cataloguing your agents, configuring the necessary gateways and identity provider, and assigning clear ownership for alerts and intervention. This guide explains what is available, what each control requires and the four decisions your team still needs to make.
What Salesforce released, and when
This table follows Salesforce's October 27 announcement. Salesforce says some features are still in development, so confirm availability before you plan around them.
Control | What it does | Availability |
Agent Kill Switch | Stops an agent's activity | Generally available now |
Triggered-Agent Kill Switch | Lets an evaluation result fire the kill switch | January 2027 |
Agent Evals | Generates tests before launch. Then evaluates agents, LLMs, and MCP servers in production against quality thresholds | Beta November 2026, general availability January 2027 |
API Evals | Builds functional and performance tests from API specs to catch breaking changes | Beta November 2026, general availability January 2027 |
Model Wallet | Sets hard limits that stop requests before costs exceed budget | Generally available now |
FinOps Dashboard | Gives one view of AI spending | Generally available now |
Cost Optimization Policies | Cuts token use with semantic caching, payload compression, and response trimming | Generally available now |
Custom Agent Scanner | Finds agents, models, MCP servers, and APIs in internal and third-party environments | Beta now, general availability January 2027 |
Open Policies for Federated API Gateways | Applies policies across Azure, Apigee, and Kong gateways | Generally available now |
The split matters. The part that stops spending and the manual kill switch work now. The part that lets a failed evaluation pull an agent offline automatically is a 2027 feature.
What each control needs before it works
Much of the coverage focuses on the feature list. MuleSoft's Get Started with Agent Fabric guide lists what you must have in place first. It also says requirements vary by entry point and by the job you do, so read it against your own setup.
Its opening advice is simple: you cannot apply governance, security, or cost controls to agents you do not know about. Catalog your assets first.
Three lines in that guide deserve a second read:
"Organizations without an IdP can't use model wallets." No identity provider, no spending caps.
Detection flags risky behavior for your review. The Kill Switch policy then blocks that agent's access to the model proxy when you quarantine it. A person still makes the call.
Agentforce agents are not A2A-compliant on their own. The guide describes an A2A bridge that presents them as compliant, with Agentforce as the supported source platform.
Plan for this as a project, not a toggle. Expect your admin, integration, and security teams to get involved.
The four decisions the tools cannot make
Salesforce describes the old problem well. After an issue surfaces, teams hunt for the owner and wait for approval while the agent keeps producing bad results and spending tokens. Agent Fabric shortens that. It does not decide anything for you.
Here are the four calls your team has to make.
At Implementology, we believe agent governance needs to be designed into the implementation from the beginning, not added after something goes wrong. Before an agent goes live, teams need to define its purpose, data access, decision boundaries, quality expectations and human escalation paths.
Agent Fabric provides controls to enforce parts of this model, but the responsibility for designing it still belongs to your team.
1. What does "good" mean for each agent?
Agent Evals scores agents against quality thresholds that you define. A vague threshold gives you a vague alarm.
Write each one in plain words first. Salesforce's own example shows why. A retailer switches its order-support agent to a cheaper model. Agent Evals runs scheduled tests and finds the agent giving wrong answers about delayed orders. That threshold has a clear test: does the answer match the order system?
Deciding what an AI agent should and should not be allowed to do is just as important as monitoring its performance. Our guide to 5 Salesforce workflows you should never hand to an AI agent explores how to establish those boundaries in practice.
2. Who gets the alert?
Name a person for every agent. Not a team and not a shared inbox.
One customer quoted in Salesforce's announcement says the value is alerting the right owner or pulling the agent offline before a small issue spreads. The tool can find the owner. You have to name one first.
3. Who can pull the plug?
The Kill Switch policy blocks an agent's access when you quarantine it, and Salesforce says approval can stay in the loop where your policy requires it. In Salesforce's example, the retailer chose to keep a human in the loop.
Decide who approves a quarantine, how fast they must respond, and who covers when they are out. A slow approval defeats the point of a fast alarm.
4. What is the spending limit, and who owns it?
MuleSoft's guide says a model wallet caps a caller's token or dollar spend against a provider. Salesforce says hard limits stop requests before costs exceed budget.
Choose which callers get a wallet, set each limit, and name a budget owner. Someone has to say yes when a team asks to raise it.
A pre-launch checklist
This list is our recommendation, built from Salesforce's announcement and MuleSoft's setup guide. Work through it before an agent touches real customers.
Inventory every agent. Register it, or set up a scanner. Name one owner for each.
Write a quality threshold for each agent in plain words.
Connect an identity provider that issues JWTs. Set agents up as service identities.
Route model calls through a model proxy on an Omni Gateway. Apply JWT validation.
Set model costs so token counts read as dollars. Then create wallets with limits.
Decide who approves a quarantine, and how fast they must respond.
Test the manual kill switch in a non-production environment.
Send alerts to a channel that has an owner. The MuleSoft Slack app can deliver them.
Put a January 2027 review on the calendar, when Agent Evals and the triggered kill switch reach general availability.
Re-read Salesforce's and MuleSoft's pages before launch. Requirements and versions change.
Frequently asked questions
What is Salesforce Agent Fabric?
MuleSoft describes Agent Fabric as the AI control plane for agents, MCP servers, and APIs across platforms. Salesforce says it helps manage and govern agents regardless of the platform that built them.
Is the Agent Kill Switch available now?
Yes. Salesforce lists the Agent Kill Switch as generally available as of October 27, 2026. The triggered version, which an evaluation result can fire, arrives in January 2027.
When do Agent Evals launch?
Salesforce lists Agent Evals and API Evals in beta in November 2026, with general availability in January 2027.
Do I need an identity provider?For model wallets and rogue-agent detection, yes. MuleSoft's guide says an identity provider that issues JWTs is required for both, and that organizations without one cannot use model wallets.
Can Agent Fabric stop AI spending automatically?
Salesforce says Model Wallet sets hard limits that stop requests before costs exceed budget. You still have to set up a model proxy, an identity provider, and the wallet limits first.
Can a failed evaluation shut an agent down on its own?
Not yet. The triggered kill switch reaches general availability in January 2027. Until then, the kill switch is the manual one, and Salesforce says approval can stay in the loop where your policy requires it.
Does Agent Fabric only work with Agentforce?
No. Salesforce describes it as a control plane across platforms. MuleSoft's guide also describes an A2A bridge that presents Agentforce agents as A2A-compliant.
Agent Fabric gives you a smoke alarm, a spending cap, and an off switch. Your team still has to say what counts as smoke, who hears the alarm, and who can use the switch.
The manual kill switch and the spending limits work now. Automatic shutdowns and Agent Evals come in 2027. Use the time to register your agents, connect your identity provider, and name your owners. That work pays off no matter how the roadmap moves.
Planning Your Agentforce Implementation?
Building reliable AI agents takes more than choosing the right platform. It requires the right use cases, permissions, data connections, decision boundaries and ongoing optimization.
Implementology helps businesses design, build and manage Salesforce Agentforce solutions that work with their existing business systems.
From initial strategy and environment setup to agent development, integrations and team enablement, we help businesses move from AI experimentation to practical implementation.
Explore our Agentforce services to see how we can help with your AI agent implementation.
Comments