Salesforce's Enterprise AI Harness: One Control Plane for Every Agent

Daniel Reyes Principal Engineer (AI/ML), YuSMP Group · AI agents and applied LLM systems for US and EU teams

The short answer

What did Salesforce actually announce?

Salesforce framed the Trusted Enterprise AI Harness as a “composable architecture” rather than a single product. It bundles six capabilities — Trusted Context, Trusted Agency, Trusted Action, Trusted Governance, Trusted Security and Trusted Models — on top of technologies the company already ships, spanning Data 360, MuleSoft, Agentforce, Salesforce Guardian and the core platform. The new piece is the AI Control Plane: one place to see, manage and control agents and AI across the organization, including agents that were not built on Salesforce.

Concretely, Salesforce says the Control Plane lets an enterprise discover and register agents, establish identity and policy, manage their lifecycle, evaluate performance, observe behavior and outcomes, and control cost. It functions as an agent registry — defining, tracking and querying the digital “workers” running across a business — and can manage MCP and large-language-model servers underneath. The harness is built “headlessly,” with capabilities exposed through MCP, APIs, Skills and Plug-ins, and Salesforce points to interoperability spanning Anthropic’s Claude, Slack, Microsoft Teams and its own Agentforce. This is, in effect, an attempt to standardize how AI agents are built, registered and supervised at enterprise scale rather than tool by tool.

The company was candid about the timeline. Many foundation technologies are available today, but the new capabilities and the unified experience are planned to begin rolling out in early fiscal FY28, which starts in February 2027. In other words, this is a stake in the ground at Dreamforce, not a switch you flip next week. Salesforce also leaned on the argument that models are converging: “The intelligence … coming from the model is ubiquitous,” one executive put it, arguing that what stays unique is an organization’s trusted enterprise context — the data, rules and permissions the harness is meant to enforce.

Why is governance becoming the product?

For two years the competition in enterprise AI was about the model and the agent framework. Salesforce is arguing that the next contest is the layer above them: who governs the agents once every team has spun up its own. The data behind the pitch is striking. A VentureBeat Intelligence survey in July 2026, covering 107 organizations with 100 or more employees, found that 85% already run two or more agent orchestration platforms, at an average of 3.1 platforms per company. That is exactly the fragmentation — separate identities, separate policies, separate cost meters, separate audit trails — that a control plane promises to collapse.

There is a defensive logic, too. If agents proliferate outside Salesforce, Salesforce risks becoming one node among many rather than the system of record. By offering a neutral-sounding plane that governs third-party agents through open interfaces like MCP, it tries to stay central even when the agents are someone else’s. Whether that neutrality holds in practice is the open question, but the framing is deliberate: governance, observability and cost control are where large buyers actually feel pain, far more than raw autocomplete speed.

The broader market signal for teams building software is that agent governance is maturing from a checkbox into a category. Expect competing control planes from the hyperscalers and the agent-framework vendors, and expect procurement to start asking how your agents are registered, identified and monitored. Designing your GenAI integrations so that every agent has a clear identity, scope and audit trail is quickly becoming table stakes rather than a nice-to-have.

What should teams be skeptical about?

The first caution is the timeline. A preview with a February 2027 general-availability target is a direction, not a dependency. Re-architecting your agent stack around a plane that is not yet generally available would be premature; the sensible read is to note the capabilities you will want and build toward them with tools you can use today.

The second is lock-in dressed as neutrality. A control plane that governs “any” agent is enormously valuable to the vendor that owns it, because governance is sticky — once your identities, policies and audit trails live there, moving is hard. Openness via MCP and APIs is encouraging, but teams should insist that agent definitions, policies and logs remain exportable and that nothing critical is expressible only in one vendor’s console.

The third is that a governance layer does not make agents safe by itself. An agent registry tells you what is running and who owns it; it does not remove the need to review high-impact actions, scope permissions tightly, and test agent behavior against adversarial inputs. Governance makes control possible and visible; the actual discipline — least privilege, human review on consequential steps, cost limits — is still yours to enforce. For regulated FinTech and healthcare teams, that discipline is not optional, control plane or not.

What it means for US & EU software teams

The practical takeaway is that the questions Salesforce is answering are the right ones, and you should be able to answer them now, whatever platform you land on. Can you list every AI agent running in your organization? Does each have an identity and a named owner? Do you know what each can access and do, and can you cap its spend? For most teams the honest answer is “not yet” — and that gap is a risk today, not in 2027.

Closing it does not require Salesforce. An agent inventory, per-agent identities, scoped permissions, action logging and cost caps can be built on the MCP-based tooling many teams already use, and they pay off immediately: fewer shadow agents, clearer audits, no month-end surprises on model spend. Crucially, they are also the exact prerequisites that make any future control plane useful. A vendor plane inherits value from agents that are already registered and observable; it cannot retrofit governance onto agents nobody catalogued. Treating this as ordinary software delivery discipline — inventory, ownership, least privilege, monitoring — is what turns the announcement into an advantage rather than a scramble.

There is a strategic read for engineering leaders as well. If governed context, not the model, becomes the differentiator, then the investment that compounds is clean, well-permissioned enterprise data and clear policies about what agents may do with it. Teams with tidy data boundaries and explicit access rules will plug into a control plane in days; teams with sprawling, ungoverned agents will spend the first quarter of any adoption just figuring out what they already have.

What to do before general availability

Treat the announcement as a prompt to get your house in order, not to wait for a 2027 product. Here is the shippable version.

1. Inventory every agent. Build a single list of the AI agents and automations running across teams, including the ones spun up outside official channels.
2. Give each an identity and an owner. No anonymous agents. A named human is accountable for what each agent does and what it can touch.
3. Scope permissions tightly. Apply least privilege to data and actions; an agent should reach only what its job requires, and consequential actions should need human review.
4. Log behavior and outcomes. Record what agents access and do so you have an audit trail today, independent of any vendor plane.
5. Cap cost per agent. Set spend limits and per-project visibility on model usage before fan-out turns into a surprise invoice.
6. Keep it portable. Express identities, policies and logs in open, exportable form (MCP, APIs) so you can adopt a control plane — Salesforce’s or a rival’s — without a rewrite.

Used well, a control plane is a sensible consolidation: one registry, one audit trail, agents that show their work. Used as an excuse to defer the basics, it is a reason ungoverned agents keep spreading for another year. The difference is entirely in the groundwork you lay before it ships.

Frequently asked questions

Sources

Salesforce — Introducing the Trusted Enterprise AI Harness, 10 September 2026 (primary source)
VentureBeat — Companies already run 3 agent platforms. Salesforce's new Enterprise AI Harness wants to govern all of them, 10 September 2026
SiliconANGLE — Salesforce introduces Enterprise AI Harness, AI Control Plane, 10 September 2026