Salesforce Enterprise AI Harness: 6-Layer Agent Architecture
On September 10, 2026, one day before Dreamforce opens in San Francisco, Salesforce announced the Trusted Enterprise AI Harness: a six-layer architecture plus a new AI Control Plane designed to become the foundational management layer for every AI agent an enterprise runs, including agents built on competing platforms.
The announcement marks Salesforce’s most explicit bid yet to own the infrastructure surrounding AI models, not just the applications on top of them. As Rohan Kumar, Salesforce’s President and Chief Platform and Engineering Officer, put it: “Models will continue to change, and intelligence will increasingly be available everywhere. What will differentiate an enterprise is the trusted, proprietary context it brings to that intelligence.”
What the Enterprise AI Harness Is (and Is Not)
An agent harness, in the conventional sense, is the software scaffold around a model that supplies tools, memory, instructions, and an execution environment. Salesforce is deliberately widening that definition.
The Trusted Enterprise AI Harness encompasses the full stack that determines what an agent knows, what it can do, and what it is allowed to touch. It is not a new agent product layered on top of Agentforce. It is the platform architecture underneath all of Salesforce’s agent products and, by design, underneath third-party agents as well.
Salesforce is building it from reorganized technologies already inside its portfolio: Data 360, Informatica, MuleSoft and Agent Fabric, Tableau, Agentforce, Salesforce Guardian, and the Salesforce Platform. Existing customers are told they can upgrade their current investments to unlock new Harness capabilities rather than replacing what they have.
The Six Trusted Capabilities
The Harness organizes into six layers, each carrying the “Trusted” prefix to signal where Salesforce draws its differentiation from model commoditization.
| Capability | What It Provides |
|---|---|
| Trusted Context | Customer data, metadata, semantics, knowledge, real-time signals, and memory, giving agents a shared understanding of the business |
| Trusted Agency | Reasoning, planning, state, memory, collaboration, and orchestration, blending open-ended AI reasoning with deterministic controls |
| Trusted Action | Secure connections to applications, APIs, workflows, tools, and business processes so agents can complete real tasks |
| Trusted Governance | Data lineage, quality controls, guardrails, and policy enforcement over the data and processes agents rely on |
| Trusted Security | Identity, permissions, privacy, data protection, and runtime security governing what agents can access and what they can do |
| Trusted Models | Intelligent model routing based on accuracy, performance, cost, and business requirements, with flexibility to switch models as the landscape shifts |
Salesforce illustrates the architecture with a deceptively simple question: “Can we fulfill this order today?” A complete answer requires customer records, contract terms, inventory availability, business rules about what “available” means for that specific customer, the workflow to reserve inventory, and policies governing whether the agent is authorized to commit the reservation. The Harness is intended to make all of that reusable context and controlled execution, not something each agent team rebuilds from scratch.
Critically, Salesforce separates open-ended reasoning from deterministic execution. An LLM might reason probabilistically about the next step, but a business cannot necessarily tolerate that uncertainty when an agent changes an order, approves access, or touches sensitive financial data. VentureBeat noted this as the central design tension the Harness is built to resolve.
An AI Control Plane for Agent Sprawl
Alongside the six capabilities, Salesforce is introducing an AI Control Plane: a central management surface where administrators can discover and register agents, establish identity and policy, manage lifecycle, evaluate performance, observe behavior, and control cost across the entire enterprise AI estate.
The Control Plane extends to third-party agents, not just Salesforce products. SiliconANGLE described the offering as giving administrators “one place to see, manage, and control agents and AI across the enterprise,” including monitoring inference costs.
Kumar, speaking to analysts, was blunt about the coming cost challenge: “You’re going to open up the agent registry, and you’re going to see a view which is organized around what the agents are costing you across. The volume of this could be a lot.”
This is a problem every enterprise AI team will recognize. As agents multiply across business functions, unmanaged agent sprawl becomes an IT governance, budget, and security crisis. The AI Control Plane is Salesforce’s answer to that inevitable pressure, and the company is positioning it as an evolution of MuleSoft’s API management tooling extended to MCP servers, LLM endpoints, and an emerging agent registry.
Open by Design: MCP, AIforce, and Third-Party Models
One of the more consequential architectural choices is Salesforce’s insistence on composability. The Harness is being built headlessly, with capabilities exposed through MCP, APIs, Skills, and Plug-ins. This allows any AI experience, including Claude, Microsoft Teams, Slack, and Agentforce, to consume Salesforce’s trusted context and governance layers without requiring a user to work inside a Salesforce interface.
Salesforce calls this strategy AIforce: making its capabilities available to any AI, agent, application, or surface. Forkast framed it plainly: “The Trusted Enterprise AI Harness is the largest platform player’s bid to own the management layer for every agent in the enterprise, including the ones that aren’t Salesforce.”
Shawn Malhotra, CTO of Rocket Mortgage, captured the buyer perspective: “AI is moving faster than any technology we’ve seen, which is exactly why composability matters so much to us. We don’t want to bet our future on one closed stack; we want the freedom to adapt as the landscape shifts.”
The Claudeforce Connection
The Harness announcement arrives two weeks after Salesforce unveiled ClaudeForce, its deep integration with Anthropic that makes Claude the default reasoning engine across Salesforce’s platform. The Harness and ClaudeForce are complementary: ClaudeForce brings Claude’s reasoning into Salesforce workflows, while the Harness supplies the context, governance, and control layer that makes those reasoning calls reliable and auditable.
Together, the two initiatives suggest Salesforce’s theory of the agentic enterprise: the model (Claude) handles open-ended reasoning, the Harness handles the enterprise environment around it, and the AI Control Plane keeps the whole system visible and manageable. For enterprises already running Agentforce and Anthropic APIs, this is the connective architecture between them.
This also matters for teams evaluating how enterprise AI agents graduate from pilot to production. The leading failure mode in agentic deployments is not model capability. It is the absence of the surrounding infrastructure: shared context, governance, security controls, and cost visibility. The Harness directly addresses all four. Book a call to see how Enera’s design connects to this emerging architecture layer.
Timeline: What Is Available Now vs. What Is Coming
There is an important nuance in today’s announcement. Many of the underlying technologies, Data 360 data unification, Tableau semantic models, MuleSoft integrations, Agentforce agents, Salesforce Guardian security, exist today inside Salesforce’s portfolio. What is new is the architecture that connects them, the “Trusted” framing, and the AI Control Plane surface that spans all of them.
The unified Harness experience and new capabilities are planned to begin rolling out in early fiscal FY28, which starts in February 2027. Pricing details have not been disclosed. Dreamforce next week (September 15 to 17) will almost certainly include additional technical depth on the roadmap.
For Salesforce customers, the practical implication is clear: existing investments in Data Cloud, MuleSoft, and Tableau become building blocks of the Harness rather than isolated products. The upgrade path is meant to be incremental, not a rip-and-replace.
Why This Matters for Enterprise AI Builders
The Enterprise AI Harness is not the only product in this category. Snowflake’s Cortex AI Gateway, Databricks Unity AI Gateway, and a growing cohort of AI governance startups are all targeting the same control-plane problem. What distinguishes Salesforce’s entry is scale: no other vendor has the installed base of CRM, ERP integration, and analytics tooling that Salesforce can now reframe as the Harness foundation.
The battleground Kumar is describing is governance of context, not governance of models. As models become commodity infrastructure, the enterprise that controls the proprietary context layer, the customer relationships, processes, business logic, and real-time signals that make an agent useful controls the economic moat. Salesforce’s bet is that it already owns that context for a significant portion of enterprise revenue data, and the Harness is the product that makes that asset natively available to any agent.
For AI-native teams evaluating their own harness architecture, the key question the Harness poses is not whether to use Salesforce. It is whether your own context, governance, and control-plane strategy is as composable, auditable, and model-agnostic as the enterprise AI market is now demanding it be.