Product1 publisher2 min readPublished
Salesforce dates the rest of its agent harness to a fiscal year starting next February
Salesforce says many of the technologies behind its new Enterprise AI Harness already run in its cloud services, and the rest arrive early in fiscal 2028. It has not said which features sit in which bucket.
The Product Desk · Product desk
What happened
- Salesforce previewed two offerings for building and managing AI agents: the Enterprise AI Harness and the AI Control Plane.
- The harness is a collection of technologies the company says is designed to improve the reliability and security of customers' AI agents.
- It will comprise six main feature collections, and Salesforce says half of them focus on the business data agents use to carry out their work.
- The technologies that are not already in Salesforce's clouds will start rolling out early in its 2028 fiscal year, which the company says begins next February.
- The AI Control Plane will let administrators centrally monitor agent performance and inference costs and help check that agents follow cybersecurity policies.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint A team scoping an agent build this quarter cannot tell from the preview whether the control it needs exists, so the defensible plan counts a named feature as present only once someone has found the toggle.
- exposure Somebody now has to decide, agent by agent, which datasets it may read. That approval queue lands on the security team that signs off on the rollout, and it does not shrink as more agents ship.
- decision Handing model choice to a router is a decision architects have to make on purpose. An agent's answer can change because the router picked the cheaper model that day.
- capability Per-agent inference cost becomes a number a budget review can put on the table. That changes who is able to argue for switching a particular agent off.
Salesforce says many of the technologies underneath the Enterprise AI Harness are already available in its cloud services [4], and it has not published which ones [21]. So an admin asked in a Monday planning meeting whether agent data lineage works yet is left guessing by the preview. In a story filed on September 10, 2026 [19], the fiscal year Salesforce points to begins in February 2027 [18].
Three of the six collections concern the business data an agent reads [7]. Trusted Context gathers data assets such as customer records and product availability information, with support for what Salesforce calls "real-time signals" [8]. A second collection covers information accuracy and includes a data lineage tool [9]. A third gives cybersecurity teams the ability to regulate which agent can use what dataset and how [10].
The other features treat the model as swappable. Intelligent model routing will send an agent request to the most suitable LLM based on parameters such as accuracy and cost [11]. Salesforce describes the setup it is aimed at: a frontier model for complex tasks, and a smaller, cheaper model behind simpler agents [17]. Because LLMs are non-deterministic, so the same prompt five times can produce five different answers, the harness will let customers pair agents with deterministic reliability controls [12]. It will also ease the job of connecting agents to external applications [13].
"Models will continue to change, and intelligence will increasingly be available everywhere," said Rohan Kumar, Salesforce's president and Chief Platform and Engineering Officer [15]. What will differentiate an enterprise, he said, is "the trusted, proprietary context it brings to that intelligence" [16].
Half the harness is data plumbing [6], and the reliability controls sit outside the model rather than inside it [12]. This harness is built for the case where the agent read the wrong record [7].
Which leaves a two-question test for each agent a team means to put in front of a customer before February. Does it depend on one of the six collections, and can an admin open a settings screen today and find the control? If they can find it, it is in the bucket Salesforce calls already available [4]. If they cannot, it belongs in the fiscal 2028 column of the plan [5], because the company has not drawn the line itself [21]. The preview also comes without a price and without a measured reliability figure [20].
What to watch
- A per-feature availability list from Salesforce would tell buyers which of the six collections they can switch on now.
- Pricing for the AI Control Plane, and whether central inference-cost monitoring is metered on top of existing Agentforce spend.
- Any measured reliability figure for agents run with the deterministic controls, which the preview does not include.