From features to fleets
For two years, Microsoft's AI pitch was an additive one: a Copilot button in Word, a Copilot panel in Teams, a Copilot pane in Windows. Useful, but bolted on. At Build 2026 on 2–3 June, the framing changed. Microsoft is no longer selling assistants that sit beside your work; it is selling the machinery to run autonomous agents through your work — and, crucially, the tools for a company to operate fleets of them safely.
The clearest signal was the language. Microsoft drew a line between "Copilots", which work alongside a person, and a new class it calls "Autopilots", which run on their own in the background. Scout is the first Autopilot: an always-on agent that joins Teams group chats and handles Outlook threads as a participant rather than a sidebar, with general availability targeted for October. But Scout is the demo, not the deal. The deal is that any enterprise can now build, register and govern agents of its own across Microsoft's stack — and that several of this week's other announcements only make sense when you read them as the supporting layers of that platform.
What an enterprise agent now is
The most important shift is conceptual. In Microsoft's 2026 model, an enterprise agent is not a chatbot. It is a managed actor with four properties a chatbot never had.
First, it has an identity. Through Microsoft Entra Agent ID, which reached general availability in April 2026, every agent gets its own governed account in the company directory — not a shared service login. Second, it has accountability: each agent identity is tied to a human "sponsor" answerable for its access and lifecycle, and if that person leaves, sponsorship transfers to their manager automatically. Third, it has constrained permissions, enforced through the same conditional-access and least-privilege rules that apply to staff, with an audit trail of every action. Fourth, it is built, not bought: companies assemble agents in Copilot Studio and Microsoft Foundry, choosing from Microsoft's own MAI models or third-party ones, and Copilot Studio's computer-using agents — able to operate software interfaces directly — became generally available the week before Build.
The stack Microsoft is selling
Put the week's announcements side by side and a deliberate, layered platform appears. Each piece answers a different question a cautious enterprise would ask before letting software act on its behalf.
| Layer | Product | What it answers | Status |
|---|---|---|---|
| Build | Copilot Studio + Microsoft Foundry | How do we create an agent and choose its model? | Generally available; updated through 2026 |
| Identity | Microsoft Entra Agent ID | Who is this agent, and who is accountable for it? | GA, April 2026 |
| Runtime | Microsoft Execution Containers (MXC) | How do we stop it touching things it shouldn't? | Insiders after Build; enterprise preview July |
| Govern | Microsoft Agent 365 | How do we secure, monitor and manage them at scale? | GA, 1 May 2026 |
| Showcase | Scout (Autopilot) | What does a finished autonomous agent look like? | Preview now; GA October 2026 |
| Edge | RTX Spark & Project Solara | Where does it run when not in the cloud? | Devices from autumn 2026; pilots underway |
The runtime layer is the one that has been missing from most agent products. Microsoft Execution Containers bake a policy-driven sandbox into Windows itself: an administrator declares exactly which files, folders and network resources an agent may reach, and the operating-system kernel enforces it at runtime regardless of what the agent tries to do. The launch partners — GitHub Copilot, NVIDIA, OpenAI, OpenClaw, Hermes and Manus — signal that Microsoft wants this to be the place other people's agents run too. Meanwhile, the RTX Spark superchip and the Project Solara device platform push the same model to the edge, so an agent can run locally on a laptop or a dedicated badge without a cloud round-trip. Several of these pieces also count as genuine technical milestones in their own right, which is why they appear on our AI breakthroughs timeline.
Why it's a strategy, not a launch
The business model is where the layers pay off. The governance plane, Microsoft Agent 365, is sold separately at USD 15 per user per month, or bundled into the new Microsoft 365 E7 "Frontier" suite at USD 99. That pricing is the tell. Microsoft is not primarily charging for a clever agent; it is charging to manage agents — to give them identities, secure them with Defender and Purview, and keep them inside policy. The more agents an organisation builds, the more it needs that plane. Industry guidance already frames Agent 365 as effectively required once a company runs more than a handful of custom agents.
This is a familiar playbook. Microsoft did not win the enterprise by having the best word processor or the best email client in isolation; it won by owning identity, management and the directory that everything else plugged into. The agent stack applies the same logic to a new unit of work. Models are increasingly commoditised — this very week brought strong open-weight releases from NVIDIA and MiniMax, and Microsoft happily lets customers run third-party models in Foundry. The durable position is not the model; it is being the layer that names, permissions, contains, bills and audits every agent an enterprise runs. Sell the governance, and the models underneath become interchangeable plumbing.
The open questions
Execution is the entire game, and several questions remain open. The first is cost stacking: between Copilot licences, Agent 365, Entra suites and the hardware, the bill for a fully governed agent estate could climb quickly, and the value will depend on agents reliably doing useful work rather than generating supervision overhead. The second is trust. Letting an autonomous agent act inside email and group chats — sending, writing, deciding — raises a higher bar than a chat assistant that only answers when asked; Microsoft's emphasis on human sign-off for sensitive actions is an acknowledgement of exactly that risk.
The third is fragmentation. Some of this week's adjacent agent features launched without UK and European availability, a reminder that the rollout will be uneven across regions and that "available now" rarely means everywhere. And there is a broader debate, voiced by some commentators, about whether always-on agents designed to become part of daily workflows are optimised for genuine productivity or simply for engagement and dependence — a question worth keeping in view as adoption grows, even if the answer will only be clear in hindsight. As we have noted before on how AI is reshaping management roles, the technology rarely lands the way the launch slides predict.
What is not in doubt is the ambition. In a single week, Microsoft moved from selling AI features to selling the operating system those features run on. Whether enterprises buy the whole stack — or pick the layers and leave the lock-in — will be one of the defining commercial questions of the next year.