Salesforce Agent Fabric: Tests, Kill Switch and Budget Caps – and Who Approves Before the Emergency Stop

Imagine giving a new employee a company credit card with no limit and then leaving for the weekend. Nobody would do that. With AI agents it happens surprisingly often: an agent runs, calls models, consumes compute, and whether it is doing good or bad work only becomes visible when someone complains or the invoice arrives.

That is the problem addressed by an announcement from Salesforce on October 27, 2026. The vendor is expanding its control center for AI agents, "Agent Fabric", with automatic tests for agents, an emergency stop ("Agent Kill Switch") and cost controls, including a "Model Wallet" with hard spending limits. This article sorts out what the primary source actually says, what is available, what arrives only as a beta, and what a mid-sized company can take from it even if it is not a Salesforce customer.

We stick strictly to the source. Where Salesforce makes vendor claims we write "according to Salesforce". Where the source says nothing, for instance on region or price, we say so. We have assigned the topic to our Memory-Layer, and the article ends with the difference that matters: an agent with limits and an approval step versus an agent that is simply allowed to run.

What Salesforce announced on October 27, 2026

The primary source is the Salesforce Newsroom post "Agent Fabric: Catch Failures, Control AI Costs", published on October 27, 2026. Four of the additions are particularly relevant for companies introducing agents.

• Agent Evals: According to Salesforce, it generates tests before an agent goes live and then evaluates agents, language models and MCP servers in production. When checks fail, predefined interventions are triggered.
• API Evals: Generates functional and performance tests from API specifications.
• Agent Kill Switch: An intervention that Agent Evals can trigger, in other words the emergency stop for an agent whose results are no longer right.
• FinOps Dashboard, Model Wallet and Cost Optimization Policies: The cost side. According to Salesforce the Model Wallet sets hard spending limits, and the Cost Optimization Policies are meant to reduce token usage.

The post gives two figures. First, Agent Fabric handles "over 1.2 billion requests daily". Second, the Cost Optimization Policies cut token usage by up to 75 percent "based on test scenarios". Both are vendor claims, and the post does not explain how the 75 percent was measured beyond the reference to test scenarios. For your own planning, read it as an upper bound under lab conditions, not an expected value.

A third statement is more interesting for mid-sized companies. Salesforce cites a LangChain finding that fewer than half of organizations with agents evaluate them on live traffic. We did not check the LangChain report ourselves, so we treat it as a pointer, not as evidence.

An agent nobody measures is not a thrifty agent, just an unobserved one. Measuring and limiting belong to the rollout, not to later optimization.

What is available today and what is still to come

With announcements the exact availability matters, because "announced" quickly sounds like "usable". According to the post of October 27, 2026:

• Generally available since October 27, 2026: the Agent Kill Switch, the FinOps Dashboard, the Model Wallet, the Cost Optimization Policies and the "Open Policies for Federated API Gateways".
• Beta from November 2026, generally available in January 2027: Agent Evals and API Evals.
• Generally available only in January 2027: the "Triggered-Agent Kill Switch", the variant triggered automatically by checks.
• Beta as of October 27, 2026, generally available in January 2027: the Custom Agent Scanner.

What the source does not say: in which regions the features are available, what they cost and which packages include them. Whether they are available to customers in Germany, Austria and Switzerland is open. DACH availability is therefore unknown.

For you this means the Evals, the automatic quality tests, are not something most companies can use today. That is what is "coming toward companies". The cost controls and the emergency stop are, according to the vendor, already there, but within a Salesforce platform. If you do not run Salesforce agents, the announcement mainly gives you a clear template for what a sensible agent setup should contain.

The core: three questions every agent rollout must answer

1. How much may the agent cost?

An agent that calls models causes running costs that cannot be predicted like a salary. In an endless loop, with unexpectedly long documents or a faulty integration it can consume far more than planned. A hard limit, as the Model Wallet offers according to Salesforce, is therefore not a finance-department toy but a safeguard. The principle transfers: before the first productive run, set a monthly amount at which the agent stops and a person is informed.

2. How do we notice that the agent is getting worse?

Agents change without anyone touching a line of code: a provider switches the model, a data format changes, a new document type appears. According to Salesforce the core of Evals is to test agents before launch and keep evaluating them in operation so that such degradation is noticed. Even without the product: define a small number of test cases whose correct answer you know and measure the agent against them regularly. A dozen cases beat none.

3. Who may stop the agent, and who must agree first?

The third question is the most important and the one most often left open. According to Salesforce, interventions including the kill switch happen "with approval where the team's policy requires it". That is an important sentence: the emergency stop intervenes before a human could react, but whether a person must agree first is decided by the team in its policy. Whether approval is required by default or only when configured is not stated. Check this before relying on the switch.

The post describes a retailer example in which the agent asks for approval to be quarantined, reflecting the team's preference to keep a human in the loop. For us this is the real lesson: an emergency stop does not replace the decision about who decides.

Why this is your topic even without Salesforce

Salesforce writes for Salesforce customers, so the product itself is not directly relevant to many mid-sized companies. The underlying problem concerns everyone introducing agents, whether with Salesforce, HubSpot, a self-built solution or a tool like n8n.

A thought experiment, not a customer report: an agent sorts incoming supplier emails at a manufacturer with 80 employees, matches them with orders and creates queries. After a quiet month people get used to it. Then an important supplier changes its invoice format and the agent assigns invoices incorrectly. Without measurement this only shows at payment approval. With a small test series checking ten known examples each morning it shows after a day. And without a cost limit, a fault in the integration that sends the agent into a loop can cause considerable cost overnight. The scenario is constructed, but each element is a known failure pattern of agents.

Assignment: the Memory-Layer as the home for limits that apply to everyone

On our website we describe the Memory-Layer as "shared context across all agents": nothing has to be explained twice. We therefore assign control and budget to the Memory-Layer. The reasoning is our own classification, not a Salesforce product statement: when several agents work in a company, rules that apply to all belong in one central place, not in every single agent. That includes a shared budget, a shared list of things no agent may do without consent, and a shared understanding of what counts as an error.

How we proceed in the audit

We do not claim projects here; we describe our approach in the audit when it comes to control and cost.

1. Collect tasks: Which recurring tasks could an agent take over, and how much time do they cost today? We calculate with your numbers.
2. Determine the damage limit: What happens in the worst case if the agent is wrong? That decides where a person must approve.
3. Set a budget: A monthly amount at which the agent stops.
4. Define test cases: A small series of known examples against which the agent is measured regularly.
5. Name the emergency stop: Who may stop the agent, and who is informed?

What you can prepare now

Because the checking features are announced as beta from November 2026 and generally available from January 2027, and DACH availability is open, we describe preparation, not a usage guide.

1. Inventory: List which AI features and agents already run in your company, including small ones set up by individual employees.
2. Cost frame: Check whether your model or platform provider offers a spending limit and set it.
3. Collect test cases: Write down ten examples for which you know the correct answer of your future agent.
4. Clarify responsibility: Name one person who may stop an agent and one who is informed.
5. Check your Salesforce contract: If you use Salesforce, ask your contact which of these features are available in your contract and region. The source gives no details.

Limits of the announcement: what we cannot confirm

• Whether 75 percent fewer tokens is reached in real operations. The source mentions test scenarios.
• What the features cost and which packages include them.
• In which regions they are available and whether DACH is included.
• Whether approval before an intervention is required by default or must be configured.
• How reliably the automatic tests detect degradation. There is no independent verification.

This is not an argument against the features but a reminder not to confuse vendor claims with measurements.

An agent with limits and an approval step

An agent without limits works as long as there is something to do and money available. It has no limit, no measurement and nobody to stop it. When it goes off track you notice on the invoice or in a complaint.

An agent with limits and an approval step has a fixed budget, is measured against known examples and has a named person who may stop it. For actions that cannot be undone, a human decides. On our website this shows in concrete places: emails are created as drafts and not sent directly, and content does not go live without a brand check by Magdalena. The Memory-Layer is where such shared rules for all agents should live.

The Salesforce announcement is above all a confirmation. That tests, an emergency stop and budget caps are now sold by a large vendor as a product shows how seriously the question is taken. You do not need to buy a large product to ask the right questions. Ask them before the first agent goes productive.

Source: Salesforce Newsroom, "Agent Fabric: Catch Failures, Control AI Costs", October 27, 2026 (primary source). The LangChain figure is taken from this post and was not checked separately.