Microsoft Wants to Govern Every AI Agent, Even the Ones That Compete With It
For the past couple of years, the question everyone asked about AI in business was about the model. Which lab is ahead, which one will own the frontier, whether a cheaper open weight model can do the job, I have spent plenty of time on that question myself. But there is another question that enterprises started asking this year, and it is less glamorous and more consequential. When an AI agent is allowed to act inside your company, who decides that it is allowed to, what it may touch, and what it is spending? For most of the way these tools have worked, that decision has been scattered, ad hoc, and personal. Someone pastes a credential into a tool and an assistant quietly gains the power to read mail and move money. The interesting turn of the past week is that a very large company has decided to build the answer to that question for everyone, including for agents that compete directly with its own.
The company is Microsoft, and the thing is its Agent 365 platform, which it expanded at Ignite this week. Microsoft's number, and it is a striking one even allowing for vendor spin, is that organizations have registered more than fifty million agents in it since the service launched in May. Inside its own environment Microsoft says it has mapped more than half a million agents that generate something like sixty five thousand responses a day for employees. The scale is less important than what the platform now does with all of that.
Agent 365 is what Microsoft calls a control plane for agents. In practice that means it gives an agent an identity in the company's directory, applies the company's access controls to it, watches what it does, and adds up what it costs. It can also discover what Microsoft calls shadow agents, the ones that people have quietly built or installed on their own laptops without telling IT. When it finds one, it registers it and governs its access rather than pretending it does not exist.
The part that deserves the most attention is that this is not limited to Microsoft's own agents. This week Microsoft said agents built on Anthropic, OpenAI, Google, Salesforce, Databricks and Snowflake now get their own identity in its directory, so that IT and security teams can govern them alongside the Microsoft ones. That is the sentence worth re-reading. The company that sells Copilot, its own assistant and its own set of agents, is offering to be the administrator for the rivals' agents too.
This is a familiar play if you remember how Microsoft won the last time. The thing that made Microsoft the center of the enterprise was not any single application. It was the directory, the layer that answers who is allowed to log in and what they are allowed to do. Once everyone's usernames, passwords and permissions lived there, it did not matter much which applications a company preferred, the seat of power had already been taken. The agent is becoming the new unit of access, and the agent's identity is the new thing that has to live somewhere. Microsoft is heading straight for the position it occupied before, the admin console that everything else has to be compatible with.
There is a reason this, rather than the model, is where the durable value is landing. The models have been commoditizing for a year and a half, and the frontier premium keeps shrinking. Anyone can now run a capable model, and companies have grown comfortable switching. But trust is not commoditizing. The permission that lets an agent read a customer record or issue a refund is a scarce and increasingly expensive thing, and someone has to own the decision about whether it is granted. That decision has to live in one place, and the company that controls that place controls the flow of agentic work through the enterprise, no matter whose model is doing the reasoning underneath. It is the classic picks and shovels position, the layer underneath the whole rush rather than any single claim to it, and it is being built at exactly the moment the intelligence above it is becoming a commodity.
The irony is hard to miss, and it is worth sitting with because it reveals the strategy. Microsoft competes with all of these agents at the application layer. It wants people to use Copilot and its own agents instead of building on someone else's. Yet here it is, eagerly promising to govern the others, to give them identities, permissions and budgets. That apparent generosity is the point. The neutral administrator is the most powerful position in this market, because neutrality is what lets one company sit underneath all of its competitors. Acting as the referee is a safer way to control the game than trying to win every play yourself.
I want to be careful not to overstate how complete all of this is. Governing an agent means different things at different depths. Some of what Microsoft is offering is inventory and visibility, knowing that an agent exists, who owns it, and how much it is spending. Full runtime control, the ability to reach inside an agent built on a rival platform and limit what it can do at the moment it acts, is much harder, and independent commentary has already noted that many third party agent workflows remain outside what Entra can truly enforce. So the fifty million number and the neutrality story are partly a claim on where the market is heading, not just a description of what works today. That does not make the direction less significant. It makes it more interesting, because a company rarely builds an entire control plane for a destination it doubts.
For a founder, the thing to notice is not only the opportunity but the new dependence. The real dispute was never only which model is best. It is who owns the layer that decides which agents get to work. Every company that deploys agents is now going to choose, or be nudged into choosing, a governance layer, and that layer is a dependency exactly like the model was, except stickier. The cost tracking is the on ramp. It is easy to let one system start counting what your agents spend and who they are, and then to discover that the same system, or the same vendor, is the only one that can turn that off. You can swap a model in a morning. Swapping the directory where every agent's identity and permissions live is a project of a different order. The moat has moved from the intelligence to the permissioning, and so has the lock in.
I do not think this means every company should rush to build its own governance, or that Microsoft is inevitable. There are open and multi vendor efforts moving in the same direction, and the outcome is genuinely uncertain. But the direction is not. Less of the value is sitting in the intelligence, and more of it is concentrating in the layer that decides which machines may act, right alongside an expanding dependence on whoever runs that layer. For anyone building on agents, the durable position is the one that owns the relationship with a customer's trust rather than the one that merely supplies the reasoning.
The models will keep getting cheaper and easier to replace. The list of which agents are allowed to act, and who keeps that list, is where the power is quietly moving.