Salesforce Introduces the Trusted Enterprise AI Harness

It’s Dreamforce week, and for a year that’s already been packed with Agentforce and security updates (plus a fair share of renames and new partnerships), people’s expectations are high for what’s coming. Last week, Salesforce posted about its most ambitious AI architecture announcement yet: the Trusted Enterprise AI Harness. When I first heard of it, I can’t deny it felt like another buzzword drop. But taking a closer look, I realized it’s not really a shiny new feature, but more like Salesforce finally naming the direction it’s been building toward all along.

This article will cover what’s in it, and what it means for the people who will eventually be tasked with implementing whatever comes out the other end.

Six Trusted Capabilities

The announcement signaled an update on the architecture side, where Salesforce introduced six “trusted capabilities” which are: Context, Agency, Action, Governance, Security, and Models. Right above all six is the AI Control Plane, which is essentially a single dashboard for seeing every agent running across your org, what it’s allowed to do, and what it’s costing you.

One example Salesforce uses to explain how the six pieces work together is this: 

A customer asks, “Can we fulfill this order today?“, but no single system holds the full answer. In short, separate entities contribute a specific part of the scenario; for example, CRM knows the customer, while ERP knows inventory, and Contracts hold the entitlements. Analytics defines what “available” means for the business context. 

The Enterprise AI Harness is Salesforce’s answer to stitching all of these separate entities together so an agent can act on it instead of simply describing it. Here’s how each capability handles a piece of that scenario:

1. Trusted Context: Pulls together the customer’s data, contract terms, and inventory so the agent understands the full picture instead of “just a slice of it” from each.
2. Trusted Agency: Reasons through the decision, figuring out whether the order can really be fulfilled and what has to happen next.
3. Trusted Action: Carries out the decision, reserving inventory, updating the order, kicking off fulfillment.
4. Trusted Governance: Makes sure the answer is based on trustworthy data and that policy indeed gets followed.
5. Trusted Security: Controls what the agent can see and touch based on identity and permissions.
6. Trusted Models: Routes the task to whichever model makes the most with regard to cost, accuracy, and performance. It can even let you swap models later without starting over.

So, What’s New?

Right now I know this all sounds like old features in “new packaging” for things that Salesforce already has, such as Data 360, Agent Fabric, Agentforce, Guardian, the core platform and so on. You’re right to think that way because it is, in fact, not new.

None of these signal new products or acquisitions, but instead a new framing. Salesforce is painting a clearer picture for consumers that its products were meant to work together, where each piece is essential to form one whole system. Now they have given it an actual name and structure: the Trusted Enterprise AI Harness. 

I, in fact, see this as a continuation of the Headless 360 direction since TDX earlier this year, as the harness is being built headlessly – proven by its capability of being reachable through MCP, APIs, Skills, and plug-ins instead of something that’s only configurable within Salesforce’s own UI. 

It’s also the reason why the announcement calls out Claude, Slack, and Microsoft Teams as surfaces this can show up in, instead of simply highlighting Agentforce as the only user-facing interface.