Your Agents Now Get Tested Before Customers

Salesforce’s Agent Fabric can now test your AI agents before customers meet them, and keep testing them once they are live. On 27 October 2026 Salesforce added Agent Evals, API Evals and a set of hard cost controls to Agent Fabric, the MuleSoft control plane that manages agents across platforms. For teams running Agentforce, or a mix of Agentforce and other vendors’ agents, it is one of the most practical announcements of the autumn. Quality and spend move from things you check after the fact to things the platform enforces while the agent works.

The timing suits most organisations well. Agent Evals and API Evals enter beta in November and reach general availability in January 2027, which leaves a clear window to prepare. The cost controls and the Agent Kill Switch are available today. A team can start capping AI spend this month and arrive in January with a test suite ready to switch on.

The teams that will get the most from this are the ones who already know what a good answer looks like in their own sales and service processes. That knowledge lives in the CRM, and it is the raw material for every evaluation the new tooling runs. Here is what shipped, what it means for cost and ROI, and how to use the next ten weeks well.

What did Salesforce add to Agent Fabric on 27 October?

Salesforce added four groups of capability to Agent Fabric on 27 October 2026. Agent Evals generates tests before an agent is deployed, then checks agents, language models and MCP servers in production against quality thresholds you set. API Evals builds functional and performance tests straight from API specifications, so a breaking change is caught before it reaches production. A FinOps Dashboard, a Model Wallet with hard spending limits and Cost Optimization Policies give budget owners one view of AI spend and a ceiling that stops requests before a budget is exceeded. The Agent Kill Switch lets a team take any agent offline, and a version triggered automatically by failing evaluations is due in January 2027.

Two smaller additions matter for larger estates. The Custom Agent Scanner, in beta now, lets customers build their own scanners to find agents, models, MCP servers and APIs behind firewalls or on third-party platforms and list them in the Agent Registry. Open Policies for Federated API Gateways applies one policy, such as rate limiting or IP allowlisting, across Azure, Apigee and Kong gateways in a single action. Salesforce says Agent Fabric already handles more than 1.2 billion requests a day. ASICS Digital is helping shape the Agent Evals beta, and Capita, GE Appliances and Newell Brands are also quoted in the announcement.

What is Salesforce Agentforce, and where does Agent Fabric fit?

Salesforce Agentforce is the platform for building and running AI agents inside Salesforce. Its agents read CRM data, follow business rules and take actions such as updating a case, qualifying a lead or drafting a quote. They reason over instructions and topics, call actions built in Flow, Apex or MuleSoft, and hand over to a person when a request falls outside their scope. Agent Fabric sits one level above. It is a control plane that discovers, governs and now tests agents wherever they were built, whether in Agentforce, another vendor’s platform or an in-house framework. Agentforce is where an agent does its work, and Agent Fabric is where you see, measure and manage every agent at once.

That separation is useful because very few organisations run agents in one place. A study Salesforce cites in the announcement puts the share of enterprises running agents on more than one platform at 96 per cent. A Salesforce service agent might sit alongside a Copilot agent in finance and a specialist support agent such as Fin. Agent Fabric gives all of them one registry, one set of policies and now one testing regime. Microsoft’s Agent 365 and ServiceNow’s AI Control Tower are pursuing the same role, which is good news for buyers: governed, measurable agents are becoming a standard expectation across the market.

How do agent evals keep AI costs on budget?

Salesforce projects, and AI work in particular, tend to go over budget when consumption is hard to see and changes are hard to test, so teams over-provision and move slowly. Futurum’s August 2026 survey of CIOs found 46.9 per cent of enterprises over budget on AI. The new controls address both causes. The Model Wallet sets a hard limit that stops requests before spend passes the budget, and the FinOps Dashboard shows consumption by agent and model in one place. Cost Optimization Policies use semantic caching, payload compression and response trimming, which Salesforce says cut token usage by up to 75 per cent in test scenarios. Evals then make the biggest saving safe: moving an agent to a cheaper model.

Salesforce’s own example shows how that works. A retailer moved its order-support agent to a less expensive model. Scheduled Agent Evals caught that the agent had started giving inaccurate information about delayed orders. With approval, the agent was quarantined, its instructions were revised and the evaluation rerun before it went back into production. Without evals, a model change like that is a leap of faith, so most teams avoid it and keep paying for the premium model. With evals, it becomes a routine optimisation that can be tried, measured and reversed quickly. The 75 per cent figure comes from test scenarios, and the beta is the ideal place to measure it against your own traffic.

Your CRM processes are already the test cases

CRM data quality is important because every AI agent built on the CRM inherits it. An agent that answers order questions is only as accurate as the order, shipment and account records it reads, and an evaluation is only as meaningful as the expected answers it compares against. Clean, consistent data gives agents correct inputs and gives evaluations reliable benchmarks. That turns data quality from a housekeeping task into the foundation of measurable AI. Teams with well-maintained opportunity stages, case categories and product data can write meaningful tests quickly, and every improvement to the data shows up directly as a higher pass rate.

A good test suite is written in business language before it is written in code. What should the agent say when a customer asks about a delayed delivery to a site in Gothenburg? Which discount may a quoting agent offer without approval? When should a service agent hand over to a person? Sales, service and RevOps teams answer these questions every day, and the answers already exist in playbooks, approval matrices and validation rules. Collecting twenty to thirty of them per agent, with the expected outcome next to each, gives Agent Evals the thresholds it needs on the first day of the beta.

How do you measure the ROI of an AI agent?

Measuring the ROI of an AI agent, like measuring CRM ROI generally, means tying the system to business outcomes rather than activity. For an agent that starts with three numbers: the cost per resolved interaction, the share of interactions resolved without a person, and the quality of those resolutions against an agreed standard. Agent Fabric now supplies two of the three directly. The FinOps Dashboard gives cost by agent and model, and Agent Evals gives a quality score over time. Combine those with resolution data from the CRM itself and you have a return figure a finance director can audit, which is exactly what wins budget for the next agent.

Measurement is still the exception, which makes it an advantage. Salesforce cites a LangChain finding that fewer than half of organisations evaluate their agents on live traffic. The teams that do can show a quality score holding steady while cost per interaction falls, the clearest case there is for widening an agent’s scope. A sensible pattern is to agree the quality threshold with the business owner before go-live, publish the score monthly next to cost, and treat any model or prompt change as an experiment with a before-and-after reading. Over a quarter, that record becomes the business case for the next deployment.

Evals turn agent sprawl into a manageable backlog

Reducing Salesforce technical debt starts with knowing what you have and what each component is for, then fixing or retiring items in order of risk and value. Agents are the newest source of that debt: prompts edited without review, actions that call APIs nobody owns, and pilots that quietly became production. The Custom Agent Scanner and Agent Registry give a complete inventory, and API Evals flag the integrations that would break on the next change. Together they turn a vague sense of sprawl into a ranked backlog. Teams can then consolidate overlapping agents, retire unused ones and document the rest, which keeps the org quick to change for every later release.

The Agent Kill Switch belongs in the same conversation, and it is a confidence tool more than an emergency one. Business owners say yes to new agents faster when they know any agent can be paused in seconds and restored once its evaluation passes again. GE Appliances described the value of a single control plane in exactly those terms. The triggered version, due in January 2027, links a failing evaluation to an automatic pause, with human approval wherever policy requires it. Setting those approval rules now, while volumes are modest, means the automated version can be switched on the day it reaches general availability.

The Sirocco perspective

We see the same pattern across Salesforce, HubSpot and Microsoft Dynamics 365: the platforms are making agents measurable, and the organisations that benefit first are the ones that can describe good work precisely. An independent CRM partner is useful here because the test cases, approval rules and cost thresholds belong to the business process, not to any one vendor’s tooling. We help clients write them once, in plain language, and apply them to whichever control plane they run, whether that is Agent Fabric, Agent 365 or both. Because we do not sell licences, our advice on where an agent should run, and which model it should use, starts from the client’s process and budget.

For Salesforce customers, the next ten weeks are a good moment to move. Switch on the Model Wallet and FinOps Dashboard now, since both are generally available. Pick one agent, write its first thirty test cases with the people who own that process, and join the Agent Evals beta in November. By January, when evals and the triggered kill switch reach general availability, you will have a tested agent, a cost baseline and a quality score to show for it, the strongest possible starting point for scaling agents across the business. If you would like help planning that, schedule a consultation with our team.

Get in Touch

Preparing an Agentforce agent for the Agent Evals beta, or setting AI cost limits across Salesforce and other platforms? Tell us where you are and we will help you write the first test suite.

So where do you start?

As your long-term partner for sustainable success, Sirocco is here to help you achieve your business goals. Contact us today to discuss your specific needs and book a free consultation or workshop to get started!