Salesforce Pitches an Enterprise AI Harness and a Control Plane to Govern Cross-Vendor Agents

Salesforce introduced a new architecture it calls the Trusted Enterprise AI Harness on September 10, five days before its annual Dreamforce conference opens in San Francisco, pairing six "trusted" capabilities with a central AI Control Plane meant to give companies one place to see, govern and pay for AI agents across their business. The company framed the announcement as the foundation of the "agentic enterprise," a way to let agents reason and act on proprietary business context while staying inside corporate guardrails. It also made clear that most of the architecture is planned rather than shipping: many of the underlying technologies already exist across Salesforce products, but the new capabilities and a unified experience are scheduled to begin rolling out in early fiscal 2028, which starts in February 2027.

Related reading: Microsoft Borrows $25 Billion as the AI Buildout Moves Onto the Bond Market · Amazon's AWS Grew 39% While Free Cash Flow Fell to -$42.4 Billion.

The timing matters. Dreamforce runs September 15 to 17 with an explicit "Agentic Enterprise" theme, so the harness is better read as a strategic marker laid down ahead of the company's biggest marketing moment than as a product customers can buy today. Salesforce itself says details on availability, packaging, pricing and upgrade paths will be announced closer to general availability, and warns that pricing and packaging are subject to change.

What the harness actually is

The word "harness" normally describes the scaffolding around a language model: the system prompt, tools, execution hooks and context management that turn a raw model into an agent that can do useful work. Salesforce is stretching that definition to cover the enterprise plumbing around agents, including the data they read, the actions they take, who is allowed to do what, and which model gets each job. The company says the resulting system can be assembled from Salesforce technology, a customer's existing tools, or both, and can incorporate third-party models, agents and systems.

The architecture is organized into six capabilities, each prefixed with "Trusted," according to Salesforce's announcement:

• Trusted Context assembles customer records, metadata, semantics, knowledge, real-time signals and memory so agents share a common understanding of the business and the customer.
• Trusted Agency covers reasoning, planning, state, memory, collaboration and orchestration, mixing flexible model reasoning with deterministic controls where a process needs certainty.
• Trusted Action connects agents to applications, APIs, workflows and tools so they can reserve inventory, update records, trigger fulfillment or hand off to a person.
• Trusted Governance adds lineage, data quality, policy enforcement and guardrails over the data and processes AI relies on.
• Trusted Security applies identity, permissions, privacy, data protection and runtime security to what agents can see and do.
• Trusted Models lets companies connect different models and route work by accuracy, performance, cost and business requirements.

Salesforce's own illustration is a deceptively simple customer question: "Can we fulfill this order today?" No single system answers it. A CRM holds the relationship, an ERP holds inventory, contracts hold entitlements, analytics defines what "available" means, and company policy decides what can be promised. The harness is meant to stitch those inputs into reusable context and controls that multiple agents and models can share.

Alongside the six capabilities sits the AI Control Plane. It is designed to let administrators discover and register agents and AI capabilities, assign identity and policy, manage lifecycles, evaluate performance, observe behavior and outcomes, and control cost, across both Salesforce and non-Salesforce AI. The pitch is consistency: a company should not have to manage every agent or AI experience separately as they proliferate across teams, applications and models.

A platform-power move, dressed as plumbing

The strategic subtext is control of the enterprise agent layer. Salesforce is arguing that the models themselves are becoming interchangeable, and that the durable advantage lies in the proprietary context a company wraps around them: its customers, transactions, processes and history. "The Agentic Enterprise won't be defined by which model a company chooses," Rohan Kumar, Salesforce's president and chief platform and engineering officer, said in the announcement. "Models will continue to change, and intelligence will increasingly be available everywhere."

VentureBeat reported that Kumar, in an interview, described model intelligence as ubiquitous and the enterprise's "trusted context" as the differentiator. In the same article, Kumar framed the harness as composable, saying customers do not have to source every layer from Salesforce. Salesforce says the harness is being built headlessly and exposed through MCP, APIs, Skills and Plug-ins, so its capabilities can extend into Claude, Slack, Microsoft Teams, Agentforce and other surfaces, part of a broader "AIforce" strategy.

That openness is a tell. If enterprises already run agents from several vendors, a control plane that can only see one vendor's agents would address only part of the problem. Salesforce is betting that the layer above the agents, where identity, policy, observability and cost live, is where lock-in and margin will accumulate. It is not a novel idea, but Salesforce is arriving with a distribution advantage: CRM, Slack, Data 360, MuleSoft and Informatica already sit inside many large companies.

The product is also a consolidation story. Salesforce folded capabilities from six different products, including MuleSoft Agent Fabric, Informatica and Data 360, into the harness. MuleSoft Agent Fabric was itself previewed at Dreamforce a year earlier and released in January 2026, touted then as an "agentic control plane." The rebranding suggests Salesforce decided the individual pieces needed a single narrative and a single console, TechTarget reported.

The competition is already shipping

Salesforce is not alone in chasing the agent control layer, and its rivals have a head start on availability. Microsoft's Agent 365 has been generally available for commercial customers since May, positioning itself as a way to observe, govern and secure agents even outside Microsoft's ecosystem. Amazon made its Bedrock AgentCore Harness generally available in June. ServiceNow offers AI Control Tower, Genesys has an AI Control Plane, and AWS runs a Multi-Provider Generative AI Gateway, with a cluster of startups pitching narrower governance tools.

That is why the "early fiscal 2028" wording matters so much. Analysts covering the space know the market is moving quickly, and VentureBeat's assessment was blunt: it is unlikely customers will wait for a branded harness when governed alternatives are purchasable now. Microsoft and AWS are also arguing for openness within their own stacks, so Salesforce cannot win on the mere claim of being multi-vendor; it has to be better at context and cost than the alternatives.

There is evidence that the market genuinely is fragmented. VentureBeat Intelligence's July 2026 Agentic Orchestration Pulse Survey, based on 107 respondents at organizations with 100 or more employees, found that 85 percent run two or more agent orchestration platforms, with an average of 3.1 per enterprise. A majority, 53 percent, expected their primary agent control plane to be hybrid by the end of 2026. VentureBeat noted the survey was self-selected and skewed toward large technology organizations, so the figures are directional rather than a market-share estimate. Still, they describe the reality Salesforce is pitching against: agent estates are already heterogeneous.

Analyst Brent Leary of CRM Essentials put the stakes plainly, telling TechTarget that every company wants to control as many integration points as possible because customers use several vendors, and that controlling cross-vendor orchestration puts a vendor "in the catbird seat." Salesforce's data-management general manager, Savinay Berry, compared the moment to the early cloud market, predicting consolidation and simplification, but faster than cloud took.

What "early FY28" means for customers

For enterprise buyers, the practical reading is that almost nothing changes this year. Salesforce says many of the technologies underpinning the harness are available now inside its products, so existing customers already own pieces of the foundation. The new capabilities and the unified experience are what is pending. Existing customers will be able to upgrade eligible Salesforce investments to unlock new capabilities as they arrive, but Salesforce has not said which editions or products qualify, what a customer must upgrade to, whether individual capabilities can be licensed separately, or how usage across models, agents, data and actions will be metered. Those answers are promised closer to general availability.

That gap between architecture and commerce is the biggest open question. A control plane's value depends on whether it is cheaper and simpler than assembling governance, observability and identity independently. Salesforce has said it wants a simple and predictable model, with customers paying for what they use and a "common currency" spanning the capabilities, but it has not published how much will be sold through conventional licensing versus consumption.

There is also a technical caveat worth flagging. Salesforce AI researchers published a preprint on September 8 that found dramatically improving an agent's harness around a smaller model could lift average task success from 29.2 percent to 78 percent without changing model weights. But when they then fine-tuned the weaker model to imitate a stronger model, success fell to 63.1 percent, which the authors attributed to a loss of model-harness fit; a more targeted approach recovered it to 79.7 percent. The study examines a narrower technical harness, not the Enterprise AI Harness, and it is a preprint rather than a production evaluation. Even so, it complicates the promise that enterprises can swap models freely: a harness tuned for one model may not transfer cleanly to another.

What to watch

Dreamforce next week should supply the missing detail. The company has promised "the biggest product launches, boldest ideas, and most exciting announcements" at the event, and the harness is likely to be the organizing idea connecting its AI agents, data platform and integrations. Watch for any concrete dates, pilot programs or named customers beyond Rocket Mortgage, whose chief technology officer, Shawn Malhotra, is quoted in the announcement saying the company does not want to "bet our future on one closed stack."

Three things would turn the strategy from framing into fact. First, a firm general-availability window and a shipping definition of the unified experience. Second, pricing that makes the control plane competitive with Microsoft and AWS rather than merely complementary to them. Third, independent evidence that Salesforce's context and governance layer works on agents built outside Salesforce, which is the claim the entire composability argument rests on.

Until then, the honest summary is that Salesforce has described a coherent architecture and a plausible strategic bet, while leaving customers to weigh it against tools that are shipping today. The gap between those two positions is exactly what the next two quarters will decide.

Cover image: the entrance to Salesforce's Dreamforce conference in San Francisco. Photo by InvadingInvader, via Wikimedia Commons, CC BY-SA 4.0.