Salesforce announced Agent Evals, API Evals and spending controls for Agent Fabric on 27 October, saying the cross-platform agent control plane handles more than 1.2 billion requests daily. The additions are designed to test agents in production, initiate interventions when quality falls below a set threshold and stop requests before they exceed a budget.
Key points
- Salesforce targets a November beta and January 2027 general availability for Agent Evals.
- A basic kill switch is available; linking evaluation results to a policy-driven shutdown is scheduled for January 2027.
- Salesforce says its budget controls can stop requests before spending exceeds a set limit.
Agent Evals moves towards production testing
Agent Fabric is Salesforce’s control plane for managing agents across platforms. Salesforce says Agent Evals will generate tests before deployment and assess agents, language models and Model Context Protocol (MCP) servers in production against quality thresholds set by a customer. A failed evaluation can initiate a predefined response, including an alert or the Agent Kill Switch. A team’s policy can require approval before the intervention proceeds.
Salesforce first outlined Agent Evals at its Dreamforce conference in September. It now targets a November 2026 beta and general availability in January 2027, VentureBeat reported. API Evals, which Salesforce says will generate functional and performance tests from the software interfaces agents depend on, has the same planned release schedule.
The basic Agent Kill Switch became generally available in September. Salesforce has scheduled the Triggered-Agent Kill Switch, which connects evaluation results to policy-driven intervention, for January 2027. The existing control blocks governed traffic through a gateway and can be applied to a request, session, agent or tenant. Blocking that traffic does not necessarily stop every process on the platform hosting an agent.
Salesforce’s example follows a retailer moving an order-support agent to a cheaper model. In the hypothetical workflow, tests find inaccurate answers about delayed orders, the system requests approval to quarantine the agent, and the team revises its instructions before testing again. Poonam Deshpande, director of engineering for digital commerce at ASICS Digital, said in Salesforce’s announcement that her company was partnering with Salesforce to help shape the Agent Evals beta.
Model Wallet sets spending limits
Salesforce says its FinOps Dashboard gives budget owners a view of AI spending, while Model Wallet applies limits that stop requests before costs pass the allocated budget. Its Cost Optimization Policies use semantic caching, payload compression and shorter responses to reduce token consumption. Salesforce says the policies cut token use by up to 75% in test scenarios.
Those controls build on work outlined over the summer rather than all arriving with Agent Evals. Andrew Comstock, Salesforce’s senior vice-president and general manager of MuleSoft and Agent Fabric, said in the company’s announcement that the product brings evaluations, spending controls and intervention into the same system. The budget owner sets the limit that governs whether further requests can proceed.
Salesforce says 96% of enterprises run agents on more than one platform, while fewer than half of organisations with agents evaluate them on live traffic. Agent Fabric is intended to manage agents built outside Salesforce as well as those on its own platform. Comstock said customers do not have to use Agentforce, Salesforce’s agent platform, to use Agent Fabric.
Custom Agent Scanner extends the registry
Salesforce also says its Custom Agent Scanner extends discovery beyond environments supported by Agent Fabric’s existing scanners. Customers can build scanners for internal systems, environments behind firewalls and third-party platforms, then catalogue agents, models, MCP servers and APIs in the Agent Registry without registering each asset individually.
Agent discovery predates the October announcement: MuleSoft described its scanners in April. Salesforce says another update, Open Policies for Federated API Gateways, is designed to apply and track policies across gateways from Azure, Apigee and Kong. The policies include controls such as rate limits and restrictions on which internet addresses may connect.
Salesforce began a beta of Custom Agent Scanner on 27 October and plans to make it generally available in January 2027, VentureBeat reported.
