Salesforce is expanding Agent Fabric with tools to test AI agents and APIs, cap spending, and respond when an agent’s quality slips. The useful shift is from spotting a problem to setting controls that can act on it, though some of the headline features are not available yet.
Salesforce Watch analysis
What happened
Salesforce says Agent Evals will test agents, language models and MCP servers before deployment and against defined quality thresholds in production. If results fall short, teams can set responses such as alerts or taking an agent offline, subject to their approval policies. Agent Evals is due in beta in November 2026, with general availability planned for January 2027.
API Evals will generate functional and performance tests from API specifications to catch breaking changes. Salesforce also says its FinOps Dashboard and Model Wallet give budget owners a view of AI spending and hard limits that stop requests before costs exceed budget. Those cost controls are generally available now. Custom Agent Scanner, which discovers agents and related assets across internal and third-party environments, is in beta; open policies for Azure, Apigee and Kong gateways are generally available.
Why it matters
A model change can save money and still make an agent worse at answering customers. Salesforce’s example is an order-support agent switched to a cheaper model: evaluations spot inaccurate answers, then the system can request approval to quarantine it while the team fixes and retests it. That connects evaluation to an operational response, rather than leaving someone to notice the problem and chase down the right owner.
The controls are aimed at organisations running agents across several platforms, where budgets, APIs and responsibility can otherwise become a game of organisational pass-the-parcel. They do not guarantee accurate outputs: teams still need useful tests, clear thresholds and sensible intervention policies.
Our read
This is a meaningful enterprise-control release, not simply another dashboard with a reassuring name. The strongest idea is linking quality checks to an agreed response, while leaving room for human approval where policy requires it. Buyers should distinguish what is available today from features still in beta or planned for 2027; the calendar, as ever, is also part of the product.
What to watch
- Whether Agent Evals reaches beta in November and general availability in January as planned.
- How teams define quality thresholds and decide when an agent should be quarantined.
- Whether the cost limits and API checks work across the mix of platforms enterprises actually use.
Discussion spark: When an AI agent fails a quality check, should the default be to pause it automatically or wait for a person to approve the intervention?
Sources and evidence
not affiliated with or endorsed by Salesforce