Salesforce used the Dreamforce 2026 window to announce something that goes beyond product features. On September 10, the company introduced the Trusted Enterprise AI Harness, a composable architecture designed to bring all of its AI infrastructure under one shared framework and give businesses a single operational layer for managing AI agents across their entire stack.

This is not a single product launch. It is Salesforce putting an architectural name on a set of six capabilities that now sit underneath everything else it sells.

The Six Capabilities

The Harness groups enterprise AI into six “trusted” dimensions: context, agency, action, governance, security, and models. Each one maps to a real operational problem businesses face when they move AI from pilot into production.

Trusted Context addresses how agents understand the customer and the business. Without a shared data layer, every agent operates with its own partial view. Salesforce builds this on its Data 360 infrastructure and the Informatica acquisition.

Trusted Agency covers how agents reason and act on behalf of the business. This is where Agentforce lives, along with the orchestration layer that coordinates multiple agents working together.

Trusted Action governs what agents are actually permitted to do. Agents connecting to MuleSoft integrations and third-party systems need clear guardrails on which actions require human approval and which can proceed automatically.

Trusted Governance is the compliance and audit layer, managing lifecycle, performance evaluation, and behavioral monitoring across agents deployed throughout the business.

Trusted Security sits across the whole stack, handled through Salesforce Guardian, which monitors for unusual agent behavior and policy violations.

Trusted Models manages the AI models themselves: which models are approved for which workflows, how model access is routed, and how costs are tracked.

The AI Control Plane

Running underneath all six capabilities is the AI Control Plane, which Salesforce describes as the operational layer for managing AI at scale. It registers agents, assigns identity and access policy, manages agent lifecycle from deployment to deprecation, evaluates performance, observes behavior, and tracks cost. Critically, it works across Salesforce systems and third-party AI from other vendors.

Salesforce says many of the technologies underpinning the Harness are already shipping in its cloud products. The full unified management experience is planned to begin rolling out in early fiscal 2028, which starts February 2027.

The timing matters. Salesforce also confirmed six of its seven new named Agentforce agents are generally available now: Casey for customer service, Paige for employee IT and HR requests, Carter for shopping assistance, Marshall, Piper, and Fin. Hunter, the seventh, is in pilot with general availability planned for November.

What This Means for Business

Salesforce is making a clear architectural bet: the next phase of enterprise AI is not about deploying more agents, it is about governing the ones you already have.

Most businesses running AI right now have an accumulation problem. Agents got deployed across departments. Vendors added AI features to existing SaaS tools without asking anyone in IT. A few internal teams built their own workflows on top of open models. The result is an AI stack that nobody has a complete picture of.

The Harness framework is Salesforce’s answer to that problem. By naming the six operational dimensions and building a control plane that surfaces all agents in one place, the company is trying to move enterprise AI from scattered experimentation toward something that can actually be managed.

A few things worth watching here:

The Harness is described as composable, which means businesses should be able to adopt pieces of it without replacing everything. That matters for enterprises that have already invested in specific components from other vendors.

The AI Control Plane’s ability to manage third-party agents, not just Agentforce ones, is the part most worth pressure-testing. If it works as described, it gives IT leaders a practical governance layer across a multi-vendor AI environment. If it requires heavy customization to connect external agents, the value proposition narrows considerably.

The February 2027 rollout timeline for the full unified experience means this is partly a preview. Businesses evaluating enterprise AI governance now should track whether the promised cross-vendor capabilities ship on schedule.

What This Means for Enterprise DNA Clients

The shift from “build AI” to “govern AI” is one of the biggest inflection points in the current enterprise technology cycle. Understanding your own AI stack well enough to govern it starts with knowing what you have deployed and how it connects.

If your business is running AI across tools, teams, and vendors without a clear governance picture, that is the conversation to start now, not after something goes wrong.

Talk to an Omni Advisor about auditing your current AI deployment and building a governance foundation that scales.