The AI agent rush is on. Trust is still catching up.
Dreamforce 2026 drew nearly 44,000 Trailblazers to San Francisco and millions more online, including CIOs from some of the world’s top enterprises. Amid the announcements and product launches, one theme dominated the conversation: how enterprises can responsibly scale from a handful of AI agents to thousands.
Agent security, identity and access control have become top priorities as agents proliferate across the enterprise. That means confronting operational questions: What access should an agent have? Should it carry the same permissions as a human employee? How do you keep a human in the loop as the volume scales from a handful of agents today to potentially thousands within months?
The risk is not hypothetical. An ungoverned agent that says the wrong thing to a customer or takes an action it should not have real business consequences.
But the hard questions are not only technical. Who is actually responsible for building these agents? How do you let the enterprise experiment without stifling the creativity that got us here? And then there is cost. How do you build the governance and education needed to keep usage sustainable so no one defaults to the most expensive model when a smaller one will do?
These are not just technology questions. CIOs are acting on three fronts at once: the technology itself, the people who run it and the processes that hold it all together.
The Dreamforce keynote unpacked the AI Harness, the layer that sits between AI models and the enterprise, connecting the two so agents can act with real business awareness. It’s built on a simple truth: models alone can’t run an enterprise. They can analyze, plan and create, but they don’t actually know your business. A model knows what an order is, but not that the customer behind it has called three times this week, or that the shipment is stuck in Italy. The AI Harness closes that gap by connecting models to real business context and history, then linking them to your agents and systems so they can act on your behalf, all while staying governed. The payoff shows up directly in the business: faster-moving agents that grow revenue, delight customers, and cut costs without IT losing control.
This shift matters for IT leaders for a reassuring reason: it means they don’t have to start over. The AI Harness are many of the things already in a company’s IT estate — its data, applications, workflows and governance policies — surrounding and supporting the AI itself. That’s why the foundation matters so much. The faster CIOs move on AI, the more that foundation counts.
There is a fourth thread running through all of this: openness. CIOs do not want to lock themselves into a single vendor for the life of their AI strategy. They want the freedom to choose the right model for the job, both today and as the landscape evolves. Salesforce and NVIDIA jointly announced Koa, Salesforce’s first CRM reasoning model built for Agentforce. It is designed to help agents reason through complex, multistep workflows and select the right tools to complete them. For CIOs, that is the point: more choice in how their AI strategy takes shape without giving up control.
The Bottom Line
Agent proliferation is no longer a future planning exercise. It is happening now and the conversation has shifted from “What can agents do?” to “How do we let thousands of them operate without losing control?”
The answer CIOs walked away with is not a roadmap for removing and replacing existing systems. It is the Enterprise AI Harness, which pairs probabilistic agents with the deterministic CRM systems, data and governance already in place, along with the freedom to choose the right model for the job as the landscape evolves.
The mandate is clear: Keep investing in that foundation, treat agent identity and access with the same rigor as employee identity and access and let CRM do what models cannot: enforce the rules the business actually runs on.
