Whatever legitimate places AI has, inside an OS ain't one

Opinion Making software would be the perfect job if it wasn't for those darn users. Windows head honcho Pavan Davuluri would be forgiven for feeling this of late as his happy online paean about Windows becoming an "agentic OS" was met by massive dissent in the comments. "Agentic schmentic, we want reliability, usability, and stability" was the gist.

Davuluri replied with a classic "we hear you," admitting there was work to be done on those points, but rather curiously claimed that there had to be a balance between what user feedback said and what his team heard through "other channels." Nicely imprecise, but capable of a precise parsing in one respect at least: reliability, usability, and stability aren't coming first. If they were, he'd say so. He doesn't. They're not.

Microsoft is in the business of engineering, and engineering has its rules. All design is compromise, and innovation means doing the unexpected. Prioritization of the fundamentals is not incompatible with either, rather, it's the foundation for both. Without that, something else is going on. That something else is summed up in a two-word phrase - agentic OS. Because whatever an agentic OS is, it's not good engineering.

An operating system has a very clear job to do, one that has evolved over time and will evolve into the future, but one that remains absolutely clear. It controls the resources, virtual or physical, of a computer, presenting them as a set of standard abstract services to the applications and environment that the user interacts with. Otherwise, it should get out of the way. True for MS-DOS 1.0, true for Windows 11. Along the way, what was a quick and dirty clone of an 8-bit OS has absorbed multiple CPU architectures, multitasking, security, and the equivalent silicon evolution of single-cell algae to a planetary ecosystem of eight billion intelligent apes. Yet it's doing the same job, albeit at breathtakingly different scales.

Agentic computing doesn't fit here. It's not an OS service. Arguably, it runs counter to core OS principles. Agentic computing is, well, about agents, things that do tasks on behalf of the user within the modern, diverse mix of local and remote services and applications. An agentic platform makes sense, in the same way that a SaaS platform makes sense. A "SaaS OS" is a nonsensical concept, at least as far as desktop computing goes.

Agentic components should live above the application layer, where the user lives. They have to acquire data, run services, and initiate actions, but in ways that respect the user.

The only reason to put agentic functionality within the OS proper is for privileged access to data and services that go beyond what the user and user-level apps need. A modern OS devotes a lot of time to keeping things compartmentalized and secure. There is nothing here that needs to provide special dispensation or powers to agentic processes that are taking on user tasks.

Calling Windows an agentic OS rather than a platform might seem like semantics, but in engineering, saying what you mean matters. Microsoft has always liked eliding the Windows OS with the complete bundle of apps and non-OS services it comes with. Is the Windows desktop environment part of the core OS? Practically, yes; architecturally, no. Linux with its multiple choices of desktop environment is much less ambiguous here, but that's done widespread adoption does no favors.

What's worse is insisting that a user-level application or feature is so intrinsic to the OS that it cannot be removed. Infamously, Bill Gates argued in the 1990s US antitrust court case that Internet Explorer was so integral to Windows that it was technically impossible for it to be removed. As court documents revealed at the time, IE was actually integral to Microsoft's determination to make the company an unremovable integral part of the web. To anyone with a workable knowledge of OS architecture back then, it was obvious how untenable the claims were. Inasmuch as the scramble for AI dominance now is as powerful as that to control the web back then, it's appropriate to apply the same critical thinking.

Back in the 1990s, people desperately wanted to get on the web. The desperation behind agentic AI is coming from elsewhere, presumably the undefined "other channels" rather than user feedback.

You have to use an OS, but you can choose to adopt a platform, and an agentic platform would have degrees of adoption so low it could liquefy helium. This is market engineering - at least, let's hope so.

Whatever legitimate places AI has, inside an OS ain't one. We'll take all the reliability, usability, and stability you've got, though. ®