It's Wednesday morning and Samar Abbas is making time for me. As co-founder and CEO of Temporal, his days are always action-packed. I can tell, however, that he has been looking forward to this conversation. We're both calling in from home, excited to start the day this way.
I've worked with Samar for less than a year, when I joined the company to focus on AI engineering. In that time I’ve come to know him as someone who is, at heart, a passionate builder.
Today, I want to talk to him about an analogy he's been using: that we're in the “MS-DOS era” of AI agents. This phrase could sound like a fluffy talking point. But for Samar, it's technically grounded and personally meaningful.
"My first interaction with a computer was in the MS-DOS era," Samar tells me. He grew up in Pakistan and came to programming relatively late, around 18. "It was so cool. You have a machine with an operating system on it, and then you take software, there's a file system. The first thing you're trying to learn is, what's the file system? What are the applications you have installed on this thing?"
He's describing the experience of a teenager working with a PC for the first time, but the words would resonate with anyone setting up an AI system today. The architecture that emerged for productionizing agents (a sandbox with a file system, memory, and I/O) maps closely to the early PC.
I ask him where the analogy came from, and he's direct about it: "That's why it felt so real for me. I saw that: Oh, we unlocked this power, we democratized computing, everyone has it. And now we are democratizing AI."
"I remember one big unlock," he says, and his voice shifts: you can tell he's not analyzing anymore, he's actually remembering. "One step function change happens when you open up a socket and you create a port and you talk to another machine. That was such an insane moment. What? I can have a program that sends messages to another machine?"
He tells me about the first distributed program he ever wrote. He was in college, living in a hostel, and he wanted to find prime numbers: a problem that demands enormous computing power, especially by the standards of the time. So he wrote software that turned his classmates' computers into workers.
"I said, okay, folks, let me write something where I use your computer as a worker, so I can distribute the computation. Suddenly we went from 'one thing is useful' to 'everything combined, we can solve powerful problems.'"
I find myself smiling at this. There's something delightful about the image: a young man in a college dorm, talking his hall mates into donating their CPU cycles so he can hunt for primes. He would later go on to build distributed systems at Microsoft and then AWS and then Uber, before he'd co-found Temporal. But this moment embodies a distilled version of these future endeavors, a group working together on a timeless problem bounded only by the available computing power.
The progression from one machine to many is the arc that Samar sees for AI agents.
"That's why this MS-DOS analogy is so real for me," he says. "Yes, we are doing useful work within a single machine. But imagine when we go further. We are not going to stop at one agent. We are going to find more and more complex problems, which require more of these agents to work coherently to solve them."
I want to know more. The jump from one machine to many machines is reflected in the modern evolution towards cloud computing. Virtualization was key to enabling cloud computing, and the technology promises to play a central role for agents too.
"This concept of a sandbox is becoming so clear," Samar says, "because these agents do all sorts of things, and you need to figure out how to limit them…there are tools. And, at the end of the day, you cannot do damage beyond that sandbox."
"I think we are going to create better and better sandboxes," he says. "That's going from MS-DOS to Windows 3.1 to Windows 95 to Windows Server." He's right, there's already fierce innovation in agent sandboxing: with low-level building blocks in the hardware and operating system being assembled in new and more powerful ways.
I can’t help but spend some time on a topic that’s particularly close to my heart: state management. “MS-DOS” expands to “Microsoft Disk Operating System.” The file system is central.
I share a humbling experience I had a few years ago when I spent months trying to design an improved version of POSIX, the most ubiquitous standard for file systems. Despite my best efforts, I kept coming back to the same decades-old design, or something close to it.
This time things will be different, says Samar. "All of the primitives that were available, which people were composing to solve problems before this AI platform shift—I feel they are going to fall apart," he says. "The scale in this new world is going to be so massive that none of those primitives are designed for it."
I agree. We are at a new point in the systems design space. Not many people have introduced entirely novel ways to think about state management, and Temporal’s work on durable execution puts it in rarefied air.
There's one more thread we pull on, the one that keeps me up at night: security.
Right now, the dominant pattern for AI agents is running them on your laptop. They potentially have access to your file system, your credentials, your everything. It's extraordinarily productive. It's also, as Samar puts it bluntly, "a ticking time bomb."
A core tension is inescapable. For agents to do meaningful work, they need access to data:your company's data, your customers' data, your internal systems. "No IT or security organization will ever allow you to just pull down secrets and store them on a developer's machine," Samar says. "Imagine an agent gets compromised, and all of your customer data is published on the dark web. That would be a company-ending event."
Today, organizations are so overwhelmed by the value they're seeing from AI that many aren't fully reckoning with the exposure. "Every organization is so overwhelmed with the value, they are not recognizing the risk," Samar says. "Your security is only as good as your weakest point. And right now, everyone running these agents on their machine is an insanely insecure thing to do."
He draws on his own history. When he joined Microsoft in 2000, the company made a decision that stunned the industry: they halted all new development across the entire organization—thousands of engineers—to focus exclusively on security. "They trained the entire developer population on threat models. Everyone was building threat models."
"I honestly feel we're running that kind of tech debt with agents," he says. "At some point, everyone will say, okay, let's take a pause and fix this."
The implication is that agents need to move off developer machines and into secured environments with proper controls. And here's where the threads converge: that move to centralized, secured infrastructure doesn't just solve a security problem. It creates a distributed systems problem. The more fine-grained your security boundaries, the more isolated your sandboxes, the more you need robust coordination across all those pieces.
There’s no question that a big shift is coming. Samar offers one lens on it that I haven’t heard before, and it sticks with me.
"In the '90s, if you ask a CEO how many servers they're running, they'll give you an exact answer. Today, if you ask me how many servers power Temporal Cloud, I don't know. It depends on the time of day, the day of the week."
He lets that land, then: "Now, if you ask a CEO how many employees they have, they'll give you an exact answer. I believe we're moving into a world where that won't be an exact answer either. How many human employees? How many agents? What time of day? The answer will keep changing."
The workforce itself becomes elastic, scaling up and down like cloud infrastructure. The tools that we built for human-scale collaboration, including great ones, simply won’t do the job.
"Even something like Google Docs," Samar says. "It's designed for humans interacting. But imagine agents coordinating with each other. Is Google Docs designed to handle that scale? Most likely not."
Throughout our conversation, Samar keeps returning to the same conviction: "Every organization on the planet is going to rewire itself in this new world." The hard problems of distributed systems don't go away when you add intelligence. They get harder. "We live in a time where we will push boundaries on creating new primitives," Samar says. "Because the new world needs to be designed for a completely different scale."
Johann Schleier-Smith is Technical Lead for AI at Temporal Technologies.
