The Layer Microsoft Didn't Build
Microsoft announced today that Windows is now an AI agent platform.
Agent Mode. Windows Agent Store. Agent 365. In San Francisco, Satya Nadella put the framing plainly: AI is no longer about responding to a prompt. It’s about running the work.
That’s infrastructure-level language. And they mean it.
What Microsoft actually built
Agents that access tools. Agents that operate across applications. Agents that take multi-step action with minimal instruction — now packaged into Windows itself, distributed at Microsoft’s scale.
The execution layer is getting real platform investment. This is the right thing to build.
At the same time, two data points came out this week that didn’t make the keynote.
McKinsey published its 2026 AI Trust Maturity Survey — approximately 500 organizations, respondents who hold direct responsibility for AI governance and investment decisions. One finding: only about 30 percent of organizations have reached meaningful maturity in agentic AI governance and controls. Not technical deployment. Governance. The oversight layer.
And security and risk concerns are now the number one barrier to scaling agentic AI — cited by nearly two-thirds of respondents. More than regulatory uncertainty. More than technical limitations.
The thing blocking scale isn’t the agent’s capability. It’s whether anyone trusts the agent to act without being watched.
Separately, Harvard Business School published research analyzing hundreds of millions of AI agent interactions. Knowledge workers are the heaviest users. Productivity is the top category — document editing, email management, workflow automation. The researchers used a specific phrase: agents are being treated as “a second brain and pair of hands.”
Their conclusion on the limits: human oversight will remain critical, especially for high-stakes and irreversible decisions.
The gap the keynote didn’t name
Read those two findings against today’s Build announcements.
Microsoft shipped the execution layer. McKinsey says the governance layer is immature in 70 percent of organizations. HBS says the oversight layer is still human.
That’s not a contradiction. That’s a description of what’s actually missing.
The execution layer is getting faster, more capable, and more deeply integrated into the OS. The layer that sits above execution — the layer that knows which tasks are actually open, what’s been decided, what’s pending, what needs the human and what doesn’t — is still entirely manual.
This is not a trust problem. Trust is a symptom. The underlying structure is a coordination problem.
An agent in Agent 365 can draft the email, book the meeting, update the spreadsheet. What it cannot do is know which email matters today, given everything that’s changed since Monday. Which meeting actually needs to happen versus which one is carrying forward an open loop nobody resolved. Which update is urgent and which can wait another week.
That context doesn’t live in any agent. It lives in the coordination layer above them — or, more accurately, it lives in the human’s head, which is why the oversight layer is still human.
A pattern that keeps repeating
Every time the execution layer gets better, this gap gets more visible.
When software automation expanded in the 2000s, the bottleneck didn’t disappear. It moved. Execution got faster. Coordination got more demanding — more things to track, more status to communicate, more humans needed to connect the pieces.
The answer was more coordination overhead: more meetings, more status updates, more project management tools, more check-ins to re-establish context.
Now the execution layer is getting better again. Much better. What Microsoft shipped today represents a step-change in what an individual knowledge worker can execute on demand.
The coordination overhead will expand proportionally. Not because the agents are bad. Because faster execution doesn’t tell anyone which tasks should be executing at all.
What the missing infrastructure does
The layer above execution doesn’t make tasks faster. It makes the right tasks happen without being asked.
It’s the system that knows: this loop has been open for six days, the person waiting on it needs an update, the decision it’s blocking has three other tasks downstream. It tracks what changed overnight, synthesizes what needs action today versus what can wait, and surfaces only what the human actually needs to decide.
Not a to-do list. Not a calendar. A coordination layer — one that processes context across tools, channels, and commitments, and tells the execution layer what to run.
What Microsoft built today is excellent and necessary. It is also exactly one layer below the thing that would resolve the governance gap McKinsey documented and the oversight burden HBS measured.
What today actually signals
Microsoft making Windows the agent platform is a meaningful infrastructure moment. It will compress the execution layer cost for millions of knowledge workers.
It also makes the coordination gap structurally impossible to ignore. When every knowledge worker has capable execution agents — running directly in the OS, connected to their tools — the only remaining bottleneck is direction. Who decides what the agents should do, in what order, based on what’s actually open.
That question is answered by a coordination layer or it is answered by the human. Right now, it’s answered by the human. That’s what the McKinsey data is measuring. That’s what the HBS research is naming.
The execution layer just got a platform.
The coordination layer still doesn’t have one.
Eliran Keren — Founder of Deeplica, building the coordination layer for knowledge work.
Sources: Microsoft Build 2026: Windows becomes the platform for AI agents · McKinsey: State of AI Trust in 2026 · HBS Working Knowledge: Who’s Adopting AI Agents
