Salesforce rolls out “Agent Evals” and new cost controls to stop AI agent failures and runaway spending

Salesforce is expanding the way enterprises govern AI agents, introducing Agent Evals along with additional “API Evals” and tighter cost controls inside its Agent Fabric platform. The move targets a growing operational problem: as organizations scale agent-based applications, failures can remain unnoticed, producing incorrect responses while token usage and other computing costs quietly accumulate.

In an Oct. 27 announcement, Salesforce said Agent Evals is designed to continuously evaluate agent performance and spending. The company positions the capability as a control layer that can operate across platforms, regardless of where an agent is built or deployed. The goal is to give CIOs and AI teams more confidence to scale production deployments, not just run proof-of-concept pilots.

Salesforce described agent failures as something that can “silently” surface, delivering wrong answers and racking up costs before anyone detects the issue. It argues that the traditional process is too manual, requiring teams to notice problems, identify who owns the agent, then seek approvals before restricting it or pulling it offline. As a result, the company says many organizations lack ongoing checks on agents using live traffic.

The company cited usage patterns to underscore the need for oversight. It said Agent Fabric already handles more than 1.2 billion requests daily and that the need for proactive controls will grow as enterprises increasingly run agents across multiple platforms. Salesforce also stated that fewer than half of organizations with agents evaluate them on live traffic, and that 96% of enterprises running agents use more than one platform.

Technically, Salesforce said the updates connect evaluation results to predefined governance actions. In practical terms, it described a workflow where, when quality slips, Agent Fabric can automatically step in by alerting the responsible owner or pulling an agent offline. Salesforce also pointed to hard spending limits that cap usage at budget, aiming to stop cost spikes from turning into runaway spend.

Salesforce executives and customers framed the initiative as an operational upgrade for enterprises that rely on AI agents for customer-facing service and other critical functions. Andrew Comstock, SVP and GM of MuleSoft and Agent Fabric, said enterprises need to handle two separate problems today: spotting wrong answers or spending spikes, and stopping them. He said Agent Fabric unifies evaluation and cost controls so teams can contain issues quickly and scale with confidence.

The company also highlighted how Agent Fabric expands its coverage. Salesforce said Agent Fabric already automatically discovers and governs agents across major AI platforms, and that the new updates extend governance further to agents running in internal and third-party environments, as well as to APIs managed on gateways from other vendors. Salesforce did not publish detailed specifications in the release text, but it emphasized the intent to reduce gaps between environments and tooling choices.

In customer quotes, Salesforce cited companies spanning retail, manufacturing, and services. ASICS Digital’s Poonam Deshpande said the value of Agent Evals is not only diagnosing problems, but also enabling Agent Fabric to intervene automatically when quality drops, before a small issue becomes widespread. Capita’s Tiina Stephens described the platform as a central AI control plane intended to help teams orchestrate agents with operational experts, while ensuring appropriate human intervention and secure deployment.

The release included additional cautions about what is available. Salesforce said the article may reference services or features that are still in development and unreleased, and advised customers to base purchase decisions on fully released capabilities. For prospective users, questions likely remain about which evaluation and cost controls will be generally available first, how quickly they can be adopted across heterogeneous agent stacks, and what telemetry and alerting workflows customers will need to integrate for best results.

Looking ahead, teams evaluating Salesforce’s approach may want to watch for expanded Beta availability details for Agent Evals, plus any documentation on how organizations can define quality thresholds, map interventions to specific owners, and set budget caps in ways that align with their procurement and finance controls. As enterprises push agents into live customer workflows, the practical test will be whether continuous evaluation reduces incidents without slowing legitimate automation and service scaling.