Salesforce Adds Agent Evals and Hard AI Spending Limits
Salesforce announced Agent Evals, API Evals and advanced cost controls for Agent Fabric, its AI control plane, so a failing agent can be pulled offline automatically and a runaway bill can be capped before it exceeds budget.
Transcript
Salesforce says it can now catch failing AI agents automatically. It can also limit their spending.
Agent Fabric will generate tests before deployment. Salesforce says it then checks live agents against quality thresholds.
Detection is not the whole feature. Salesforce says failing agents can be pulled offline automatically, with approval where policy requires.
Hard limits now stop requests before costs pass budget. Salesforce says policies cut token use up to seventy five percent in its own testing.
The two evaluation products are beta next month, generally available in January twenty twenty seven. Salesforce published this itself, and each quote is a customer or an executive.
Inboxsmith builds AI phone agents for small business. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Every narrated claim traces to this single first-party Salesforce newsroom article, published October 27, 2026. No video transcripts were available for this story, so the article snapshot saved at scripts/data/production/salesforce-agent-fabric-evals/article.txt is the complete factual record behind the script, and nothing in the narration comes from any other source. HOOK, 'Salesforce says it can now catch failing AI agents automatically. It can also limit their spending.': supported by the article's opening paragraph, which states that 'Today, Salesforce announced Agent Evals, API Evals, and advanced cost controls for Agent Fabric, its AI control plane for managing and governing agents across platforms', that 'As enterprises scale AI, agents can fail silently, delivering wrong answers and racking up costs before anyone notices', and that 'CIOs and AI leaders can now catch problems before they turn into customer complaints or runaway costs, with automated interventions when quality slips and hard spending limits that cap usage at budget.' Both halves of the hook are Salesforce describing capabilities of its own products, which is why the hook is attributed in narration with 'Salesforce says' rather than asserted as a measured fact; the detection half in particular rests on products that are not yet generally available, as scene four states. SCENE 1, 'Agent Evals generates tests before deployment. Salesforce says it then checks live agents against quality thresholds.': supported verbatim by the Agent Evals bullet, 'Automatically generate tests before deployment, then evaluate agents, LLMs, and MCP servers in production against defined quality thresholds', and reinforced by the article's framing that 'Catching them proactively means building, maintaining, and evaluating performance continuously.' The narration deliberately says 'live agents' rather than enumerating agents, large language models and MCP servers, because the list of three would be an enumeration; nothing is added. This is a product description of an unreleased feature, not a measured result, and the second sentence is attributed to Salesforce. SCENE 2, 'Detection is not the whole feature. Salesforce says failing agents can be pulled offline automatically, with approval where policy requires.': supported by the Agent Evals bullet, 'When results fall short, they trigger predefined interventions, including the Agent Kill Switch, with approval where the team's policy requires it', and by the worked example in the article, in which 'Agent Evals automatically flags the quality issue and requests approval to quarantine the agent, based on their preference to keep a human in the loop. Once approved, the agent is quarantined while the team refines the agent's instructions'. The phrase 'pulled offline' is taken from the customer quote by Poonam Deshpande, Director of Engineering, Digital Commerce at ASICS Digital, who says Agent Fabric can act 'alerting the right owner or pulling the agent offline before a small issue becomes a widespread customer problem'; that customer is describing the Agent Evals Beta she is helping shape, so it is a vendor and customer characterisation of unreleased behaviour, not an independent test, and the narration attributes it to Salesforce. SCENE 3, 'Hard limits now stop requests before costs pass budget. Salesforce says policies cut token use up to seventy five percent in its own testing.': the first sentence is supported by the advanced cost controls bullet, 'The FinOps Dashboard and Model Wallet give budget owners one view of AI spending and hard limits that stop requests before costs exceed budget', and the word 'now' is accurate because the availability section lists 'FinOps Dashboard, Model Wallet, and Cost Optimization Policies: Generally available today.' The second sentence is supported by the same bullet, 'while Cost Optimization Policies cut token usage by up to 75% (based on test scenarios) through semantic caching, payload compression, and response trimming.' The seventy five percent figure is explicitly Salesforce's own parenthetical '(based on test scenarios)', so it is a vendor test-scenario number and not an independently measured customer outcome; the narration therefore says 'Salesforce says' and 'in its own testing', and the figure is spoken exactly once in the whole script. The narration omits the three named optimisation techniques to avoid an enumeration. SCENE 4, 'The two evaluation products are beta next month, generally available in January twenty twenty seven. Salesforce published this itself, and each quote is a customer or an executive.': the first sentence is supported directly by the availability section, which states 'Agent Evals: Beta in November 2026. Generally available January 2027.' and 'API Evals: Beta in November 2026. Generally available January 2027.'; 'the two evaluation products' refers to Agent Evals and API Evals, the subjects of scenes one and two, and 'next month' is accurate relative to the October 27, 2026 publication date. The article also carries Salesforce's standard disclaimer that 'This article may include references to services or features that are still in development and are unreleased. Customers should make their purchase decision based on fully released and available features', which is the basis for treating the evaluation capabilities throughout the script as vendor claims rather than shipped, verified behaviour. The second sentence is supported by inspection of the article itself: it is published on Salesforce's own newsroom at salesforce.com/news/stories, and every person quoted is either a Salesforce executive, namely Andrew Comstock, SVP and GM of MuleSoft and Agent Fabric, or an employee of a Salesforce customer, namely Poonam Deshpande of ASICS Digital, Tiina Stephens of Capita, Adam Jones of GE Appliances and Steve Sigrist of Newell Brands. There is no independent analyst, no third-party benchmark and no named external tester anywhere in the article. For completeness, the article contains further claims the script does not narrate and therefore does not need to support: that Agent Fabric handles 'over 1.2 billion requests' daily, that 'fewer than half of organizations with agents evaluate them on live traffic', that '96% of enterprises running agents on more than one platform', and the Custom Agent Scanner and Open Policies for Federated API Gateways extensions. The script also deliberately avoids re-announcing that Agent Fabric exists or explaining what an AI control plane is, because our channel already published Shorts on the Agent Fabric control plane launch and on Salesforce AI autonomy and approval rules; the news carried here is confined to the new evaluation products, the automatic intervention, and the hard cost limits.(Salesforce's own newsroom article, dated October 27, 2026, titled 'Salesforce Expands Agent Fabric to Automatically Stop Runaway Costs and AI Agent Failures Before They Impact Customers'.)