Hybrid AI Agents, Orchestration, and the Real Reason Microsoft is Fixing Windows 11 ⭐

I spent much of the previous weekend catching up on Build 2026 session videos, with an emphasis on those related in some way to Windows 11. In doing so, a few themes emerged, but one stood out above all others.

The sessions I watched fall into three general categories:

Windows app development. The smallest body of content, of course, but a topic Microsoft has only sporadically addressed in recent years. This was the strongest showing for Windows app development at Build in a long time.

Developer productivity. Microsoft has pushed Windows as the best place for developers for years, even if the work they’re doing is for the web, the cloud, or cross-platform. And while this has been a theme for a while now, Build 2026 was still notable for the long list of advances that Microsoft is making in this area.

Agentic AI. This was, by far, the biggest topic this year at Build, and that’s true even if you restrict the conversation just to Windows. But there is a confluence of factors that makes 2026 particularly notable for agentic AI, among them the escalating cost of cloud-based AI, the ever-improving nature of local AI, and the need to orchestrate AI between the two.

That last category is, to me, what Build 2026 was really about. We knew that agents were coming to Windows 11. But we didn’t know that they would be hybrid agents with local AI sub-agents. And that Microsoft would expand the notion of hybrid AI to mean so much more than most of us ever imagined.

This is fascinating for all kinds of reasons, key among them being that all this work has also led to the “pain points” fixes we’re seeing this year. But the seeds for this shift were planted years ago, back before Copilot was even a gleam in some marketer’s eye and Microsoft’s AI ambitions were limited to machine learning (ML) capabilities that ran only on lackluster Snapdragon-based Windows 10 PCs that no one was using anyway.

I bet you forgot about at least half of this.

From Always Connected PCs to Copilot+ PCs

Microsoft marketing is an oxymoron on the level of military intelligence, but it’s interesting to look back on the initial announcement about the Windows team partnering with Qualcomm to bring Windows 10 on Arm-based PCs to market. At that time, these new Windows 10 on Arm PCs were marketed as Always Connected PCs because of their integrated cellular modems, a key feature of the Arm-based smartphones and other mobile devices. Naturally, Intel had to get in on the action, so the Always Connected PC moniker came to select x86-based PCs too.

What does that remind you of?

Several years ago, Microsoft and Qualcomm came to market with a new initiative based on a new generation of Nuvia-powered chips for Windows 11 on Arm. This time, they went to market with Copilot+ PC, and the focus this time was on the powerful NPU integrated into these new SoCs and how it would enable incredible on-device AI experiences. Naturally, Intel had to get in on the action, so the Copilot+ PC moniker likewise came to x86-based PCs, too.

In both cases, Microsoft buried the lede. The real reasons to choose Arm-based PCs over their more familiar x86 counterparts are multifold, but it started with efficiency and battery life in 2017, and as we moved into the Snapdragon X era two years ago, that grew to include reliability, performance, and compatibility. Where the first Snapdragon-based PCs were deeply compromised, Snapdragon X was nothing short of a miracle when it first appeared in 2024.

But think about the intent behind all this. Terry Myerson, who led the Windows organization during the Windows 10 era and announced the partnership with Qualcomm on Snapdragon chips for PCs, once told me that he just wanted to “get those f#$king Intel stickers off of PCs.”

That’s a great line, but it was always more nuanced than that. It wasn’t really about stickers, literally, but about what those omnipresent logos reminded Myerson of every time he saw one: The microprocessor giant had ignored Microsoft’s increasingly frantic need for more efficient mobile processors for laptops for over a decade. And if all that Snapdragon accomplished was to get Intel to wake up to the realities of the changing market, then he was OK with that. All he really wanted was for PCs to enjoy the best attributes of mobile computing, meaning instant-on, reliable, and efficient performance with excellent battery life and, yes, always-on connectivity.

Put as simply as possible, this was one of many efforts on Microsoft’s part over many years to make the PC better. And, go figure, Microsoft did communicate that when it first announced Windows 10 on Arm and its Qualcomm partnership. Unlike the previous attempt at an Arm-based platform in Windows RT, Windows 10 on Arm would enable, “for the first time ever, … the Windows our customers know with all the apps, peripherals, and enterprise capabilities they require, on a truly mobile, power efficient, always-connected cellular PC.” And unlike with Windows RT, Windows 10 on Arm could run “x86 Win32 and universal Windows apps, including Adobe Photoshop, Microsoft Office, and popular Windows games.” That would later expand to x64 and, with Snapdragon X, this platform finally made sense for the first time ever. But whatever. Better, in this case, meant combining the best of the PC with the best of mobile. A hybrid approach, if you will.

Initiatives like this can raise all boats. Always Connected PC and Copilot+ PC were both industry-wide pushes that would see PC makers adopt the chipsets and make new thin and light laptops, often premium products, they could sell to customers. Component makers would need to come on board with drivers and utilities. Microsoft engaged with developers to help them port their apps to Arm in native form. And though this was and is directly competitive with more established x86 chipsets from Intel and AMD, Microsoft’s hope was always that it would inspire them to do better, too. As they did, slowly at first, in adopting many of the architectural benefits of Arm in their own chips, with multiple cores, multiple core types, and more efficient hybrid designs.

These things can be delicate from a relationship perspective, but Intel, in particular, was complacent and failing in ways we’d only come to understand more broadly years down the road. One might effectively argue that Microsoft forcing the issue was exactly the push that Intel needed. And that the industry benefitted fairly broadly as a result. Including even Qualcomm’s competitors.

From NPUs to hybrid AI

Though the 2024 Copilot+ PC launch was an obvious milestone on the road to the hybrid AI capabilities that will soon be mainstream, it didn’t come out of nowhere. Microsoft had been talking up machine learning (ML) and AI for years ahead of the current era, and it delivered its Windows Studio Effects suite of webcam (and audio) capabilities to NPU-equipped Snapdragon PCs in the Windows 10 era. Qualcomm, for its part, first pushed hybrid AI on PCs in May 2023, one year before the Copilot+ PC launch and the availability of its revolutionary Snapdragon X chipsets. And the industry had been using the term AI PC to describe any PC with an NPU, including the relatively tame pre-Copilot+ PC versions.

But Copilot+ PC and the 40 TOPS NPU that the specification requires were, if anything, ahead of the curve. They promised transformative capabilities but delivered just a series of mostly minor enhancements that many have correctly pointed out would work fine–if less efficiently–on normal PCs using a GPU or even a CPU. In short, there is no killer app and most of the improvements are invisible to users.

That makes for a tough sell, and it didn’t help that the NPU requirement also felt artificial, like Microsoft was trying to force new PC purchases on its users. Worse, it created a bizarre chicken and the egg situation in which potential customers didn’t see the benefits of buying a Copilot+ PC for local AI, while developers only adopted these capabilities slowly because the audience of users with Copilot+ PCs was small compared to the broader Windows 11 user base. A non-virtuous cycle, basically.

Two years into the Copilot+ PC era, second-generation Snapdragon X2-based PCs are rolling out, but slowly, and even more sporadically than was the case with Snapdragon X in 2024. PC makers still ship far more hardware based on Intel and AMD chipsets, including Copilot+ PCs. And though Microsoft has expanded the unique capabilities available on this platform, the most notable feature being Click to Do, it’s still just a scattered collection of features that feel like things that should work to some degree on any modern PC.

Concurrent to all this, AI in the cloud has exploded, and not just in the obvious functional ways. After subsidizing the cost of these expensive cloud-based AI interactions for three years and spending several hundred billion dollars collectively on AI infrastructure build-outs, Big Tech is finally done being charitable: With usage-based pricing now expanding across the industry like a virulent cancer, we’re finally being confronted with the true cost of this technology. Like this year’s gasoline prices, costs are through the roof, and they’re incommensurate with the value delivered.

Ironically, the local AI that Microsoft’s been pushing for two years could help solve this problem if it just worked well. But for various factors–the chicken/egg issue noted above, the NPU requirement, the lack of a killer app, and so on–it doesn’t work well. But what does that even mean, to work well?

Broadly speaking, the problem here is that Microsoft hasn’t been meeting its customers where they are. Yes, an NPU is dramatically more efficient than a CPU or GPU, and it can speed specific ML and AI tasks without impacting battery life or general system performance. But all its users have a PC with a CPU and at least an integrated GPU, and many have dedicated GPUs. With modern Intel, AMD, and Qualcomm chipsets–and with the coming Nvidia Spark chipsets–those non-NPU components are notably good and capable of running local AI workloads effortlessly. No one with a powerful PC cares about efficiency above all else, they just want things to work.

When we discuss hybrid AI, we typically mean some combination of local AI and cloud-based AI. This can go both ways. You may have a local AI solution that hands off to cloud-based AI only when it has to, and you can conversely fall back to local AI when you reach some time-based usage cap with cloud-based AI.

Those scenarios are well understood, but I feel like there’s more nuance here. That is, hybrid AI isn’t just about transitioning between local AI and cloud AI. To be truly viable, hybrid AI needs to orchestrate AI workload traffic intelligently based on the capabilities of the PC or other device you’re using and your access to free and paid cloud-based AI of all kinds. It needs to choose between the CPU, GPU, and NPU on your PC/device and whatever you have in the cloud.

Today, we have AI PCs that have whatever kinds of NPUs, and we have Copilot+ PCs with 40+ TOPS NPUs. Soon, we will have Nvidia-based AI PCs that are also Copilot+ PCs (in that they meet the specifications and have a powerful NPU), but those PCs (workstations, really) will also have incredibly powerful integrated GPUs. Whether there will be some new brand or name for this class of PCs remains to be seen. But what’s already clear is that Copilot+ PC will change. It will either disappear entirely or its specifications will expand to include these new Nvidia-based PCs and their powerful GPUs (and, we hope, any PC with whatever GPUs).

From Build 2026 to building a better foundation for Windows

We know all this not because Microsoft communicates effectively but because bits and pieces of the coming change, whatever it is, came to light during the Build 2026 conference last week. Among the many, many announcements were the following:

• Microsoft already said that Windows 11 would include AI agents. But now we know that Windows 11 will also include local AI agents, meaning agents that run on-device and not in the cloud. It will also include something I’m calling hybrid agents, meaning agents with cloud and local components where specific local models will power what Microsoft calls sub-agents.
• Microsoft is expanding the capabilities of its local AI models, called small language models (SLMs). Microsoft Phi is being phased out for the more powerful Aion family of SLMs, which, among other things, is what will enable local agents in Windows 11.
• Microsoft is finally expanding the Windows AI APIs for developers to work across CPUs and GPUs, and not just NPUs. This starts with speech-to-text recognition, text intelligence, and video super resolution capabilities, all of which will work on-device across some combination of CPU, GPU, and NPU without requiring a “cloud round trip.”
• The Nvidia chipsets also announced last week will be the first to run true LLMs–large language models that today run only in the cloud–locally. These local models support up to 1 trillion parameters, which was previously unthinkable. If Apple was marketing this, it would describe this as a “datacenter on your lap.” But these capabilities will of course come down to more mainstream PCs in time. And not that much time, given how fast AI evolves.

Is it ironic that Microsoft is describing these local AI capabilities as “unmetered” intelligence? Perhaps. But it’s also a good description and one that anyone paying AI usage overage fees each month will appreciate. In this hybrid AI world, LLMs–what Microsoft now calls frontier models–will tackle frontier problems while everything else (its words) will run locally and at scale. But again, to me, the orchestration is what really matters.

So let me expand on that again.

There’s this notion of hybrid AI that includes local AI and cloud AI components.

There’s this notion of an orchestrator that can determine the CPU, GPU, and NPU capabilities of the PC it’s running on and route AI workloads accordingly.

Finally, there’s also this other notion in which the orchestrator must also assess the capabilities of local AI models, perhaps installed or not, and then match them with the appropriate CPU, GPU, and NPU resources or hand off when and as needed to the cloud. Which, while always the most powerful option, may be a paid, usage-based offering with spending limits to consider.

That’s a lot of orchestration. But it’s also the key to hybrid AI and the hybrid AI agents that will scurry around on our behalf getting work done and crossing the local/cloud AI barrier as needed (securely, hopefully, thanks to Microsoft Execution Containers).

One more thing. This is, I think, a key reason why Microsoft is suddenly addressing the “pain points” that its users have complained about for years to no avail. In my evolving view of understanding the true intent of the changes Microsoft has announced, I’m starting to believe that this is the reason for the changes.

Think about it. Hybrid AI requires a solid local (PC-based) foundation, but the enshittification of Windows 11 had left its desktop platform in a sorry state with problems both real (chaotic and unreliable Windows Updates) and imagined (the made up issues with Recall, which was still a PR disaster). All the security, reliability, resource usage, and other improvements it’s made and announced over the past year or more are all about addressing those fundamental problems. No, not for you and me and our silly complaints, which don’t matter to Microsoft in the slightest. But for hybrid AI, something that matters to Microsoft quite a bit.

It’s just a theory, but like any good theory, it’s grounded (sorry) in truth. And the best news is that it doesn’t matter why Microsoft is fixing Windows 11. What matters is that it is doing that work and that all of us will benefit from it.