New AI harnesses — the code and instructions that control a generative AI model's outputs automatically, steering AI agents to complete set tasks in a more predictable fashion — are becoming almost as numerous as AI models themselves.
So it should probably come as little surprise that Salesforce, the global leader in customer relationship management (CRM) software and a fervent "AI-pilled" company, is also planning to launch one of its own: the Trusted Enterprise AI Harness, previewed today ahead of the company's annual large conference, Dreamforce, next week in San Francisco (September 15-17), but which will not be fully available to all customers until early next year.
Salesforce's Harness bundles together six capabilities — Trusted Context, Trusted Agency, Trusted Action, Trusted Governance, Trusted Security and Trusted Models — alongside a new AI Control Plane for managing agents and AI across an organization.
The strategy lands in an enterprise market that is already decidedly multi-platform. VentureBeat Intelligence’s July 2026 Agentic Orchestration Pulse Survey, based on 107 qualified respondents at organizations with 100 or more employees, found that 85% run two or more agent orchestration platforms simultaneously, with an average of 3.1 platforms per enterprise.
A total of 53% expect their primary agent control plane by the end of 2026 to be hybrid — combining provider-native and external orchestration.
The self-selected survey skewed toward large technology organizations, making the figures directional rather than a market-share estimate, but they quantify the practitioner reality behind Salesforce’s pitch: enterprises are already operating agents across heterogeneous stacks and expect the control layer above them to remain heterogeneous as well.
The move amounts to a broader argument about where enterprise AI is heading. As foundation models improve and become more widely available, Salesforce contends that the durable advantage for enterprises will come from the proprietary context and operational controls they place around them.
“The intelligence that [is] coming from the model is ubiquitous,” Rohan Kumar, Salesforce’s president and chief platform and engineering officer, said in an interview with VentureBeat ahead of the announcement. What remains unique, he argued, is an organization's “enterprise trusted context”: its customers, employees, transactions, data relationships, knowledge and history.
For enterprises, that distinction matters because increasingly autonomous agents need more than access to a capable LLM. They need to understand what a company means by terms such as revenue, customer churn or account health; which information they are allowed to access; what actions they can take; and how those actions should be monitored, governed and paid for.
“You may have the best of the model, but you won't be able to reliably reason and act on behalf of your enterprise,” Kumar said without that context.
Salesforce expands the meaning of an AI ‘harness’
Agent harnesses commonly encompass the software surrounding a model that supplies tools, context, instructions and an execution environment. Salesforce is deliberately making its definition broader.
Salesforce is borrowing a term now widely used around coding agents and expanding its boundary. Tools such as Anthropic’s Claude Code, which Anthropic describes as an agentic coding tool, DeepSeek Harness, a local-first coding-agent development and runtime environment, and OpenAI’s Symphony, an intentionally minimal orchestration layer for coding agents, largely operate closer to the execution and orchestration layer.
Salesforce wants its Harness to extend outward to encompass the enterprise context, security, governance, models and centralized controls that determine what agents know and what they are allowed to do.
Salesforce’s own AI researchers have also recently produced evidence that the narrower, technical harness around a model can substantially affect agent performance.
In a September 8 Salesforce AI preprint research paper, the company's in-house researchers define the harness as the system prompt, tool set, execution hooks and context-management scaffolding surrounding an LLM.
Across seven enterprise agent benchmarks, evolving that harness around a smaller Qwen model increased average task success from 29.2% to 78.0%, a 48.8 percentage-point improvement without changing the underlying model weights during that step.
The research also points to a complication for the plug-and-play model story. After optimizing the harness, researchers fine-tuned the weaker model to imitate trajectories produced by a stronger expert model. Instead of improving further, average success fell from 78.0% to 63.1%, with performance declining across all seven tasks.
Researchers attributed the regression to a loss of model–harness fit: the weaker model adopted the expert’s planning strategy without enough capability to execute that strategy reliably inside a harness optimized around its previous behavior. A more targeted on-policy approach, in which the expert corrected individual failing turns while preserving the weaker model’s planning trajectory, raised average success to 79.7%.
The work is a preprint rather than a production evaluation of Salesforce’s new Enterprise AI Harness, and it studies a considerably narrower technical definition of a harness. But it provides empirical support for one part of Kumar’s argument: changing what surrounds a model can dramatically change what an agent accomplishes, while also showing that the model and surrounding architecture cannot always be treated as entirely independent components.
Six layers connect reasoning to deterministic enterprise execution
Salesforce organizes its broader architecture into six “trusted” capabilities.
Trusted Context combines customer information with data, metadata, semantics, knowledge, real-time signals and memory.
Trusted Agency handles reasoning, planning, state, memory, collaboration and orchestration, while allowing deterministic controls where business processes require certainty.
Trusted Action connects agents to applications, APIs, workflows and tools.
Trusted Governance adds lineage, data quality, policies and guardrails
Trusted Security covers identity, permissions, privacy, data protection and runtime security.
Finally, Trusted Models is intended to let enterprises connect different models and route work based on factors including accuracy, performance, cost and business requirements.
Salesforce says customers will be able to use all six as one system — or they may mix and match only the components they need, combining Salesforce technology with existing and third-party systems.
The model–harness research adds an important caveat to that model-routing ambition. Salesforce wants companies to retain the flexibility to switch intelligence as models evolve, but its researchers’ findings suggest that a harness optimized around one model’s planning and execution behavior may sometimes need to evolve when that model changes.
The distinction between open-ended reasoning and deterministic enterprise execution is central to Salesforce’s pitch. An LLM might reason probabilistically about what should happen next, but a business cannot necessarily tolerate similar uncertainty when an agent reserves inventory, changes an order, approves access or touches sensitive customer information.
Salesforce illustrates the problem with a deceptively simple question: “Can we fulfill this order today?” Answering it might require CRM data about the customer, ERP inventory information, contract terms, analytics definitions, company policies, prior interactions and the workflows required to complete fulfillment. The Harness is intended to turn those separate inputs into reusable context and controls for multiple agents and models.
An AI control plane for agent sprawl
Alongside those six capabilities, Salesforce is introducing an AI Control Plane designed to give enterprises one place to discover and register agents, establish identities and policies, manage their lifecycles, evaluate performance, observe behavior and outcomes, and control cost — including across third-party AI.
That ambition lines up particularly closely with the VentureBeat Intelligence finding that enterprises already average 3.1 orchestration platforms and mostly expect control to remain hybrid. A control plane capable only of seeing agents from one vendor would address only part of the environment many respondents say they are actually operating.
It also puts Salesforce into an increasingly important competitive battle over who manages the growing population of enterprise agents.
Microsoft Agent 365, generally available for commercial customers since May, similarly positions itself as an enterprise control plane for observing, governing and securing agents, including agents operating beyond Microsoft’s own ecosystem.
AWS, meanwhile, made its Amazon Bedrock AgentCore Harness generally available in June. AWS’s definition is closer to the conventional runtime harness: it manages the orchestration loop, tools, context window, persistent state and failure recovery, but AgentCore also surrounds that harness with identity, memory, gateways, observability, evaluations and policy controls. AWS explicitly supports changing models without rebuilding agent logic.
Salesforce's differentiation is that it is trying to wrap that control-plane and runtime problem into a broader architecture spanning the context agents consume, how they reason, the actions they execute, model choice, security and governance.
Kumar said Salesforce arrived at that approach by examining where customer deployments were actually getting blocked.
“Is it cost effective? Is it secure enough? Is it producing the right results?” he said. Those questions, rather than simply whether an agent could run, increasingly determined whether enterprises trusted it with meaningful work.
Data 360, Tableau, Informatica and MuleSoft become parts of the same AI architecture
Salesforce is not building the Harness entirely from scratch. Instead, it is reorganizing technologies accumulated across its platform into a common architecture.
The company says the foundation draws from Data 360, Informatica, MuleSoft and Agent Fabric, Tableau, Agentforce, Salesforce Guardian and the Salesforce Platform.
For existing customers, current investments are intended to form the starting point, with an upgrade path to additional Harness capabilities as they become available.
The combination also shows why Salesforce is defining “context” as more than retrieving records for an LLM.
Informatica can help discover, classify and govern data. Tableau semantic models encode the vocabulary a business uses to describe metrics and operations. Data 360 brings together customer and enterprise data, including information held outside Salesforce. MuleSoft and Agent Fabric can connect agents with actions and systems.
Kumar said Salesforce is “pivoting [MuleSoft] pretty aggressively” as part of the Harness toward the control-plane problem, including understanding LLM endpoints, MCP servers and agents operating across an enterprise.
Salesforce says the stack will remain composable
Perhaps the most consequential part of the strategy is Salesforce's insistence that enterprises will not have to source every layer from Salesforce.
“You don't have to pick everything from Salesforce,” Kumar said.
The company says the Harness is being built headlessly, with capabilities exposed through MCP, APIs, Skills and Plug-ins. Salesforce specifically names Claude, Slack, Microsoft Teams and Agentforce among the experiences into which its capabilities can extend, part of what it calls a broader “AIforce” strategy to make Salesforce available to different AI systems, agents, applications and surfaces.
That openness is especially relevant against the multi-platform reality identified by VentureBeat Intelligence. Enterprises in the July survey were not converging on a single orchestration environment; most were accumulating several. Salesforce therefore has an incentive to make its context, governance and control layers useful even when the agent itself was built somewhere else.
Rocket Mortgage CTO Shawn Malhotra reinforced the openness argument in Salesforce's announcement, saying the company does not want to “bet our future on one closed stack” as AI technology changes.
That composability will be important to prove in practice. Enterprises already have models, data platforms, security systems and agent frameworks from multiple vendors, meaning Salesforce's Harness will be more compelling if it can govern and enrich those investments rather than simply becoming another stack customers have to consolidate onto.
Pricing and licensing remain unknown
For enterprise buyers, one of the biggest unanswered questions is what adopting the Harness will actually cost.
Salesforce has not announced specific SKUs, per-user or consumption rates, minimum commitments, or a definitive licensing model for the new capabilities. Its public announcement says existing customers will be able to upgrade eligible Salesforce investments to unlock additional Harness capabilities as they become available, with detailed pricing, packaging and upgrade paths to come closer to general availability.
That wording is important. Salesforce is arguing that products customers already use provide the architectural foundation for the Harness, but it has not said that future Harness capabilities will automatically be included in existing subscriptions.
Kumar said the company is still working through the commercial model. He described the guiding principles as simplicity and predictability, with customers paying according to what they use and a “common currency” spanning the capabilities. But Salesforce has not yet determined publicly how much will be sold through conventional licensing versus consumption-based pricing.
That leaves several total-cost-of-ownership questions unanswered: which current Salesforce editions and products count as eligible investments, what customers must upgrade to, whether separate Harness capabilities can be licensed independently, and how usage across models, agents, data and actions will ultimately translate into a bill.
For now, enterprises evaluating the architecture should therefore distinguish between technologies already available in their Salesforce estates and the additional Harness capabilities that Salesforce plans to introduce. The company’s promise of a simpler, composable architecture is clearer than its commercial mechanics.
Rollout begins in early fiscal FY28 — five months away
Many of the underlying Salesforce technologies are already available, but it will be a while before the full Salesforce AI Harness is available to all.
The company says new Harness capabilities and a unified experience are planned to begin rolling out in early fiscal FY28 — which begins in February 2027, according to Salesforce's calendar, a substantial amount of time given the AI industry's fast-moving, ever changing, frequently upending pace.
It's highly unlikely, in this author's assessment, that Salesforce customers or potential customers will wait around for its proprietary or branded harness in the meantime, and while they may be enticed to switch once it is fully available, it will need to be especially compelling from a price, ease-of-use, or capability perspective to make moving over to it worthwhile.
Existing customers will be able to upgrade eligible Salesforce investments as new capabilities become available. Salesforce has not yet disclosed detailed packaging, pricing or specific upgrade paths, saying those will arrive closer to general availability. Availability may also vary by region.
Those details will matter. The value proposition depends not only on whether Salesforce can technically connect context, governance, security and agent execution, but whether doing so is simpler and economically more predictable than assembling those layers independently.
Salesforce is betting that as foundation models improve, the enterprise AI battle will increasingly move outward from the model itself. The differentiator will be the architecture that can give those models proprietary business context, connect them safely to real workflows and constrain their actions without eliminating the flexibility that makes generative AI useful in the first place.
The company is not alone in concluding that the infrastructure surrounding models is becoming strategically important. What distinguishes Salesforce’s bet is its attempt to join the technical harness and enterprise control-plane concepts to the customer data, semantics, workflows and governance already embedded in its applications — while simultaneously arguing that the resulting architecture must operate across a multi-vendor AI estate.
Or, as Kumar puts it in Salesforce's announcement: “The Agentic Enterprise won’t be defined by which model a company chooses.”