Salesforce puts a control plane above the harnesses that actually run its agents

A harness, in the sense used by people who build agents, is the loop around the model. It assembles context, calls a tool, reads the result, decides whether to go again. Salesforce's use of the word covers something larger. Inside its own natural language coding tool, Agentforce Vibes, the local agent execution loops run on Mastra and the Claude Agent SDK ^[7]. The Enterprise AI Harness sits above that layer: a control plane where developers view, manage and control agents ^[8], with the six capabilities attached. The list under that label carries eight product names, and it reaches six if MuleSoft and Agent Fabric count as one capability and the Salesforce platform underneath is not a capability at all ^[14].

The framing from the top of the company puts model choice out of scope. "The Agentic Enterprise won't be defined by which model a company chooses. Models will continue to change, and intelligence will increasingly be available everywhere," said Rohan Kumar, Salesforce president and chief platform and engineering officer, during a press briefing ^[5]. What differentiates an enterprise, he said, is "the trusted, proprietary context it brings to that intelligence" ^[6]. That diagnosis matches what breaks in practice in my experience: the join across systems of record, not the quality of the completion. What follows from it commercially is consolidation of that join.

Salesforce says engineers can "use the six together as one system or take only what they need", and build with Salesforce technology, other third-party technology, or both ^[9]. That is the claim worth testing, because it is the difference between a control plane and a suite. Testing it needs interface documentation: what a non-Salesforce agent calls to read shared context, and what it calls to have a policy evaluated before an action fires. The reported launch material does not include interface documentation, licensing units or pricing ^[15].

The adoption cost lands in the data plane. Consistent governance is a property of the path an action takes, so it holds for invocations that route through the control plane and not for the ones that go around it. Buying the promise means putting customer context in Data 360 and integration governance in Informatica, both of which are inside the same bundle ^[2].

The complaint already on record is narrower. Reviewing Agentforce as a standalone tool in August, operations associate Ashish B. wrote that "One area that could be improved is the initial setup and configuration process" ^[12], and added: "Building effective agents can take some customization and a solid understanding of the workflow" ^[13]. Salesforce's answer to configuration difficulty is a larger integrated surface with a management layer over it ^[10]. As reported, the launch has no performance or adoption figures to size that trade ^[15].