A few days ago, Davuluri shared his excitement about it on his official X handle. He seemed very eager to reveal what the company has in mind at the upcoming Ignite event regarding the agentic OS plans.
Unfortunately for Microsoft and Davuluri, the response has been overwhelmingly negative, so much so that the comments on that X post have now been disabled.
Made me laugh. :)
Sorry for being an idiot, but what is an agentic OS?
Whatever it is, it sounds fucking stupid.
The OS doesnt need to be a focus. the OS is best when you completely forget its existence and can just do things without worry.
Which is why Windows 7 is the best operating system microsoft has, and seemingly will, ever produce.
“Agentic” is the buzzword to distinguish “LLM will tell you how to do it” versus “LLM will just execute the commands it thinks are right”.
Particularly if a process is GUI driven, Agentic is seen as a more theoretically useful approach since a LLM ‘how-to’ would still be tedious to walk through yourself.
Given how LLM usually mis-predicts and doesn’t do what I want, I’m no where near the point where I’d trust “Agentic” approaches. Hypothetically if it could be constrained to a domain where it can’t do anything that can’t trivially be undone, maybe, but given for example a recent VS Code issue where it turned out the “jail” placed around Agentic operations turned out to be ineffective, I’m not thinking too much of such claimed mitigations.
Wild guess:
Log everything the user is doing. Have clippy interface prompt the user to take some work off their hands. Do some web searches, start storing a dossier about the ‘project’. Give the user a rough outline to complete their project based on a trained llm and some web searches. Ask the user if the outline looks good. Ask the user if they’d like some help completing some of the steps. Burning through tokens the whole time, storing telemetry with 100% knowledge of what they user does/wants to do. Selling that exact data to project management software companies and companies that write middleware to do this work. Bind everything together into a virtual notebook where users can return to any content at any step.
This is really just a guess but… I think “agent” in this context means a personalised AI.
Training gen AI models requires huge amounts of resources. Its not practical to train an AI for your personal use.
Creating an agent is something like, taking an existing model, asking it to keep your entire browser history in mind while you ask it to do your homework.
IMO its actually one of the big limitations of gen AI, but somehow the word is supposed to mean the opposite. As in, the current approach has reached a dead end requiring exponentially more resources for less and less improvement. So because we can’t make a model that just knows or learns everything, we have to make agents that know lots about specific things.
Agentic OS is a buzzword that’s meant to imply that the OS is (or has) an AI agent doing useful things for you in the background without you explicitly asking it to do those things (ie an agent working for you). For an agent to be useful, (they say) it has to know and learn everything it can about you, your life, your friends, activities, contacts, work, and so on.
The tradeoff is pretty extreme though. Everything you do on the PC is watched, analyzed, catalogued, and retained by MIcrosoft (and possibly whoever they choose to share the info with, which is likely every government that asks). The features that do this are generically called client-side scanning and Microsoft has a few specific variants you can read about called Copilot Recall, and Copilot Vision.
Basically, system wide Clippy that is somehow even more annoying.
Look at the positive side, you can cook sausages on your overheating laptop
You mean cooking breakfast on the smoldering corpse of your laptop.
And in the non-Windows world - this could all be implemented on Linux in such a way that you, the end user choose if you want it at all, which models you want (including both local and online ones), and what gets what access. Plus all kinds of customizability. I’m sure someone is already working on this too.
big tech doesn’t have a monopoly on linear algebra or calculus.
you can run your own telemetry, analytics, and modeling pipelines with existing toolings and most modern PCs are plenty capable of doing a wide variety of useful things with said data.
these tools are very powerful. algorithmic feeds regularly hijack your dopamine response whether you’re aware of it or not. these organizations colonize your mind. your consent or acknowledgment is not a necessary factor. why should MS, Google, and Meta be the sole masters of these feedback loops?
the old privacy is dead. going into the future people should be pragmatic and actually do something about the current state of our society instead of being whiny do-nothing pissbabies complaining endlessly about things they barely comprehend. the key to breaking our chains is understanding the very tools that oppress us all.
being as tech-illiterate, as philosophically-illiterate as most people are is why we’re in the precarious position we’re in… one where these big tech empires have risen to institutional scales, rivaling nation-states themselves. and that includes the average software developer, sysadmin, IT support, doctor, engineer etc. not even considering genAI or agentic tools - we’re in a new zeitgeist where solving highly-dimensional problems is in extreme demand with very, very few people actually educated enough to do it. we don’t need more experts, those are a dime a dozen. we need renaissance men.
do you people not see what is at stake here?
you’re a boiling frog, all of you, all of us.
one day you will wake up a citizen of meta and purchase your rations with facebook scrip, and you won’t even know it’s happened.
the decisions we all collectively take now over the next 5-10 years determine whether free society survives or if we descend into a new dark age of neo-serfdom and techno-feudalism.
may god have mercy on us all.
We should kill all the big tech CEOs and investors
physical revulsion
As far as I can see MS is planning the agent not to stay in the background as you described but to be an active means to control the operating system and software functions, so that many tasks can be instructed in natural language.
They should call this version of the OS “Doors” then, as it just shows you what doors it wants, and closes it’s door in your face if you don’t stump up your precious data.
It’s some AI crap they try to push on us.
Windows 10 had a better kernel than 7. Unfortunately, that kernel was packaged with the rest of windows 10.
Win 7 sucked too.
Win XP was the beginning of enshitification.
You just have nostalgia.
Agent work can be pretty magical. I’ve been using cursor recently and the fact that it can just execute commands on your PC means you can just tell it to do something and it does. Troubleshooting as it goes