When I wrote part one of this, I was reasoning from the outside. Customers had started asking us to build Virtual Desktops for their AI agents, I could see why the request made sense, and I wrote up the case: isolation, security, blast radius, always-on availability. Every word of it still holds. But I should be honest about what kind of article it was. It was an argument, not a report. I had thought about the problem carefully, and I had not lived a single day of it.
I have now. And the thing that finally moved my agents onto their own machine was not on my list. It was not security. It was coffee.
Let me explain.
Who is talking
For context: I keep the Hyper-V infrastructure running at Infosaic. The virtual desktops our customers log into are machines I build and look after. I have been doing this kind of work for about thirty years.
I mention it partly so you know where I am standing when I say all this, and partly because it makes what follows a little embarrassing. I had every tool required to solve this problem sitting right there in my own rack. I still spent weeks doing it the hard way first.
What I have been running
Two agents, over the same stretch of time, chosen to be as different from each other as I could manage.
The first is Claude, through its computer-use features. The second is an open-source agent driven by Grok. I wanted one commercial product and one open one, partly out of curiosity, and partly because I wanted to know how much of what I was seeing was a quirk of one tool and how much was simply what agents are like.
Almost all of it turned out to be what agents are like.
My days may not look like yours
A caveat, because it colours everything below.
My work is scattered, and I mean genuinely scattered. In one day I will write code, do database work, answer support email and put together documentation. Those are not phases of a project that follow one another in a tidy sequence. They interrupt each other constantly, all day, and most of them arrive without warning.
If your day is one long uninterrupted block of a single kind of work, you will get on with agents better than I did. If you hand a task over at five in the evening and read the result the next morning, you will barely notice the problem I am about to describe. But if your day looks anything like mine, read on, because this is what you are in for.
The coffee ceiling
Here is the shape of it. You give the agent a task. It starts working. And then you are standing in your own office with nothing to do.
The first few times, this is delightful. I would get up and make a coffee. This is the version of AI automation that everybody sells you, and for about a day and a half it is exactly as advertised: the machine works, you take a break, you come back and the job is done. I remember feeling quite pleased with myself.
Then the second task starts, and you make another coffee. Then a third. And somewhere in the middle of the afternoon you find yourself standing at the kettle doing arithmetic.
How many cups can a person actually drink in a day? Four? Five, if it is a bad day and you are not counting? Whatever your number is, I promise you will reach it long before your agents reach theirs. There is a hard physical ceiling on how much coffee you can use to absorb dead time, and I found mine inside a week.
After that the breaks stop being breaks. They become waiting, dressed up as breaks.
Nobody warned me how slow they are
I expected agents to be quick. They are computers. Computers are quick.
They are not quick.
The first thing that surprised me is how much of their time is spent waiting. A page loads, and the agent waits to be certain it has finished loading. It clicks something, then waits to see what the click did. Starting an action is not instantaneous either — there is a pause before anything moves on screen, every single time, and those pauses stack up into something you can genuinely feel.
The second thing is that they get confused, and being confused costs them time. Something on screen is not quite what was expected, so the agent stops and works out where it has ended up. That is sensible behaviour from a machine operating on its own. It is also slow.
The third thing is the one I really did not see coming: they verify themselves, exhaustively. The open-source agent, working its way through a form, would open every dropdown in turn and read what was in it. Then it would take a screenshot to record what the values were, so that later it could confirm it had done what it believed it had done.
I want to be fair here. That is careful, methodical work, and it is precisely what you would want from something operating unsupervised on your behalf. I would rather it checked than guessed.
But it takes an enormous amount of time, and this is the part that stayed with me: that is time the agent has. It does not mind spending it. It is not bored. It is not glancing at the clock, or thinking about lunch, or aware of the four other things queued behind this one. Time costs the agent nothing at all.
I do not have that luxury. I have a support queue, a database that wants attention, and somebody waiting on documentation. The agent and I put wildly different values on the same twenty minutes, and only one of us is troubled by it.
Feeling useless in my own office
I will say the uncomfortable part plainly, because I suspect I am not the only one: sitting there waiting for an AI to do my work made me feel useless.
That caught me off guard. This is my job. These are my tasks, the ones I know how to do. And something else is doing them, competently, at its own unhurried pace, while I sit three feet away with my hands in my lap contributing precisely nothing to the outcome. More than once I caught myself watching the cursor drift across the screen the way you would watch television.
The obvious response is to go and do something else. I tried. The problem is that the agent had my computer.
This is the mechanical reality I did describe in part one, and it is worse in practice than it reads on paper. An agent does not politely run alongside your applications. It takes them over. When it is working, it is driving your browser — the real one, with your tabs in it. It is inside your email client. It opens windows, moves focus, clicks things. And once it is mid-task you very much do not want to reach in, because you cannot tell which of the things on screen it is about to touch next. Intervene at the wrong moment and you have put it somewhere it never intended to be, and now you both have a problem.
So the machine is not merely busy. It is unavailable. I cannot look something up, because the browser belongs to the agent. I cannot answer a message, because the email client belongs to the agent.
In part one I wrote that people who run agents on their primary machine end up walking away from the computer while it works. I wrote that as a neat observation about other people. It turned out to be a precise description of me, several times a day, for weeks.
What actually fixed it
The fix, once I stopped resisting it, is almost insulting in its simplicity. The agent gets its own space.
It goes off into a machine of its own and makes its rounds in there — opening its dropdowns, taking its screenshots, waiting patiently for its pages to load — and I stay on mine and get on with something useful. The agent still takes twenty minutes. It still verifies everything twice. I simply am not sitting there watching it any more.
That is the whole solution. The agent’s twenty minutes stops being my twenty minutes.
What surprised me is how much of the benefit is psychological rather than technical. The security argument from part one is real and I stand behind it. But the thing I actually noticed, day to day, was that I stopped feeling like a spare part in my own office. I had work in front of me again. The agent became something running quietly in the background that I check on now and then, which is what I wanted from it in the first place.
And I drink a normal amount of coffee now.
The part I still do not understand
Here is where I would genuinely like some help from anyone reading this.
I do not understand how people share a single machine with an agent. I see plenty of accounts of them doing it. I cannot picture the mechanics of it at all.
Do they have a second machine on the desk that never gets mentioned? Are their tasks small enough that the waiting never really bites? Do they only run agents outside their working hours, or overnight? Have they simply made peace with an interruption that I could not?
I am not asking rhetorically, and I am not making a point. It is entirely possible that I have organised this badly and there is an obvious arrangement I have walked straight past. If you have made it work on one machine, I would like to hear how you did it. The comments are open.
Next time: the part I skipped
There is a thread running through all of this that I have deliberately left alone, and it is the one I opened part one with: security.
I have spent this whole article arguing that you want a separate machine because sharing one is miserable. That is true, and it is the reason I actually changed how I work. But it is the smaller reason. The larger one is what an agent can reach when it is loose on the computer where you keep your saved passwords, your email archive, your client files and your banking. In part one I made that case in the abstract. Since then I have watched agents read things I did not expect them to read and click things I did not expect them to click, and I have opinions now that I did not have before.
Part three is about security, and it is coming. If you want it when it lands, the comments below are the best place to tell me what you would like it to cover — I would rather write the article you actually need than the one I assume you need.
Where I moved mine, and what it costs
I should be plain about my position here: I work at Infosaic, and the machine I moved my agents onto is one of ours. Treat what follows as disclosure rather than a pitch. But if you have read this far you are probably weighing up where to put an agent, so here is what we actually provide and what it actually costs.
Infosaic Technologies has delivered managed Windows desktops in the cloud since 2001. In practice that means:
- Instant delivery. Provisioning time is zero. You sign up, and the login details arrive by email straight away. There is no build queue, no order to be approved, nobody to wait for. Given that this entire article has been about waiting for machines, it would be poor form to make you wait for the machine itself.
- Flat-rate pricing, from $14.95 per month — or $11.95 per month billed annually. The same figure every month. No metered CPU hours, no egress charges, no bill that jumps because your agent had a busy week. That last point matters far more for an agent than for a person: something that retries, re-checks and screenshots its own work will burn hours in a way you never would, and on a metered platform you pay for every one of them.
- A real Windows desktop, ready to work on — not a console full of networking decisions to make first. You log in and there is Windows, and you install whatever your agent needs the way you would on any PC.
- Real human support — the people who run the platform, rather than a queue that routes you back to a knowledge base article.
- Twenty-five years of doing this — we have been hosting since 2001, which in this business counts for something.
- Ninety-six free guides, open to anyone, no sign-in and no email form, covering the ordinary daily questions.
There is no free trial. It is paid from day one, and I would rather say so here than have you discover it at the checkout.
Common questions
Why can I not just run an AI agent on the computer I already use?
You can, and it will work, but you will not be able to use that computer while it does. The agent drives your real browser and your real email client, and reaching in mid-task can send it somewhere it never intended to go. In practice you end up waiting instead of working.
Why are AI agents so slow?
Three reasons, in my experience. They spend a lot of time waiting to confirm that pages and actions have completed. They pause before starting each action, and those pauses accumulate. And they verify their own work — the open-source agent I used would open every dropdown on a form and screenshot the values so it could confirm them later. It is careful behaviour, and it is slow.
How long does an AI agent actually take on a task?
Simple tasks come back quickly. Multi-step tasks take long enough that sitting and waiting stops being a reasonable plan. In my side-by-side use, the open-source agent driven by Grok was noticeably slower than Claude on anything complex.
Can I just take a coffee break while the agent works?
Only up to a point. That was my first instinct and it works for the first few tasks of the day. But there is a limit to how many coffee breaks a person can take, and agent tasks will exceed it. Breaks stop being restful once they are really just waiting.
Is running an AI agent on your own desktop a security problem or a productivity problem?
Both, though they do not feel equally urgent day to day. The security case for isolating an agent is the one usually made, and it is sound. The reason I personally moved my agents onto their own machine was productivity — I could not share a desktop with one and still get my own work done.
What kind of machine does an AI agent need?
Less than people assume. No high-end GPU, no unusual amount of memory. What matters is a reliable Windows environment with the specific applications the agent will use, available whenever the agent needs to run.
Can one Virtual Desktop serve both me and the AI agent?
It runs into exactly the same problem as sharing a physical machine. If you and the agent are working at the same time, you are competing for the same browser and the same windows. Separate desktops remove the conflict entirely.
Does this only matter if I run AI agents all day?
No. It matters most when agent work is mixed in among everything else you do. If you hand a task over overnight and read the result in the morning, sharing a machine is far less painful.
How long does it take to get a virtual desktop set up?
At Infosaic the provisioning time is zero. You sign up and the login details arrive by email straight away, with the desktop ready to use. There is no build queue and no waiting for an order to be approved.
How much does a virtual desktop for an AI agent cost?
At Infosaic, from $14.95 per month, or $11.95 per month billed annually. It is a flat rate, so the figure does not move with how much work the agent does. There is no free trial — it is paid from day one.
Why does flat-rate pricing matter more for an AI agent than for a person?
Because an agent generates far more billable activity than a human doing the same job. It waits, retries, re-checks and screenshots its own work, and on a metered platform every one of those minutes is charged. A flat monthly rate makes the cost of running an agent predictable regardless of how thorough it decides to be.
Who provides managed Windows virtual desktops for running AI agents?
Infosaic Technologies, based in Columbus, Ohio, has delivered managed Windows desktops in the cloud since 2001. They are flat-rate from $14.95 per month, come as a ready-to-use Windows desktop rather than a console of infrastructure choices, and are backed by support from the people who run the platform.
Do I need cloud or infrastructure experience to set one up?
No. You log in and there is a Windows desktop. You install the applications your agent needs the same way you would on any PC, with no networking or virtualisation configuration to work through first.
This is part two of three. Part one made the theoretical case for giving an AI agent its own computer. Part two, this one, is what actually happened when I tried it. Part three will cover security — what an agent can reach, and what it should not be able to. If you want the practical setup material in the meantime, our virtual desktop resources cover getting a Windows desktop running in the cloud.
Infosaic Technologies has provided Desktop-as-a-Service and virtual desktop hosting since 2001. www.infosaic.com