Founders are pulling all-nighters to keep up with AI agents running 24/7

When the agents never sleep, neither does the founder

Aditya Sharma, 27, co-founder of Keel, a company that builds software for troubleshooting IT issues, recently pulled an all-nighter with his generative AI bots and didn't get to bed until 6 a.m. His agents do more than pull data and write code — they're building features for new customers, running research for the company's in-house models, and keeping an eye on marketing campaigns. The catch is that whenever an agent gets stuck mid-task, it needs a human to approve something or supply more information. That's why Sharma told The Wall Street Journal he's given up a regular sleep schedule just to stay on call around the clock. "The cost of an agent being stuck for eight hours is too high," he said. Until recently, he didn't even have a way to check on his agents remotely from his phone or smartwatch.

Why do AI agents demand round-the-clock attention?

As our glossary puts it, an AI agent isn't a chatbot that just answers questions — it's AI that goes out, searches, creates files, and even handles bookings on its own. Because agents plan and execute multiple steps independently, they can finish a task or hit a wall while their human overseer is asleep, and either way, someone has to sign off before the agent can move to the next step. On top of that, the models powering these agents keep improving on a month-to-month basis, forcing founders to constantly rethink how they structure their own workdays.

A day in the life of four founders

Founder	Company	Work pattern	Notes
Aditya Sharma, 27	Keel	All-nighter, bed at 6 a.m.	Started using a smartwatch to check on agents remotely
Peter Pezaris, 56	Proxon	7:30 a.m. to 2 a.m.	Estimates roughly 30x throughput; projects $1M+ annual revenue by October
Abby Grills, 33	Riveter	10 hours a day, one day off a week	Just two co-founders; using agents instead of hiring
Ajay Kalia, 43	Alt	5:30 a.m. to 9-10 p.m.	Controls agents via Apple Watch, gets buzzed for approvals every 10 minutes

Pezaris, who's spent more than 30 years in tech and successfully sold four companies, had promised his wife he'd retire this year. Earlier in his career, overwork landed him in the hospital twice — once after a stretch where he slept just one hour a night for a month straight and collapsed at the office. Despite that history, he launched his fifth startup, Proxon, four months ago. With a team of six engineers, the company estimates it now moves about 30 times faster than it would without agents. But there's a catch: as agents generate more output, they generate more human work too — faster onboarding, for instance, means more customer support requests coming in.

Riveter, a data collection platform co-founded by Grills and Cody Watters, still runs with just two people. They've talked about hiring, but every time the question comes up, they find a way to hand the new workload to an agent instead — and hiring keeps getting pushed back. Grills went through Y Combinator's accelerator program in 2024, where she watched a fellow founder in her cohort end up hospitalized after struggling to eat solid food and relying on meal replacements to get by. That experience is why she now makes a point of actually resting when she takes time off. Y Combinator has long encouraged its founders to take care of themselves, for what it's worth.

Kalia, who runs Alt out of Boston, says he sensed AI was about to upend everything back in February 2023, while working at Spotify's AI group. He watched colleagues with PhDs and 20 years of experience say things like "my PhD is worthless now." He left Spotify in 2024 to start Alt, and he's recently started managing his agents through his Apple Watch — which buzzes with approval requests every 10 minutes, meaning he has to check his wrist even in the middle of a run.

The FOMO of "missing a week's worth of work"

Pezaris compares working with agents to being on drugs — though he admits he's never actually tried drugs, he figures this is probably what it feels like. Underneath that feeling is a kind of FOMO: every moment he's not working feels, to him, like losing a week's worth of progress. Sharma admits he's gained weight over the past few months and finds it harder than before to spend time with people outside of work hours. Kalia says that within his network of founders, nobody talks openly about rest or burnout, but plenty of them seem to be running on fumes. Everyone quietly knows the pace isn't sustainable, he says — they just haven't found a reason to stop.

The broader shift toward delegating work to agents

It's worth revisiting our earlier coverage: One in five US workers now delegates tasks to AI instead of colleagues.

That survey, conducted by Epoch AI and Ipsos last July among 1,106 U.S. workers, found that 20% said they now hand off tasks to AI that used to go to coworkers or outside contractors. AI adoption rates varied sharply by task type, from 25% for record-keeping to 57% for systems and software design. As founders offload more and more work to agents, this WSJ reporting suggests something isn't shrinking to match: the human burden of monitoring, approving, and handling exceptions.

Editor's take

What this reporting really shows is that AI agents haven't turned out to be a labor-saving tool — they've become a new workforce that needs supervising. We're not yet at the point where a person just gives instructions and collects results. Right now, the structure demands that a human step in the instant an agent gets stuck, which means founders' working hours haven't shrunk at all — if anything, they've expanded into round-the-clock standby duty.

The contrast with the old startup "9-9-6" culture (9 a.m. to 9 p.m., six days a week) makes this clearer. Back then, the limit on output was simply how many hours a person could work. Now, agents finish tasks faster than humans ever could, so the new bottleneck is founders failing to keep pace with their own agents. Pezaris's estimate of "30 times faster" isn't an exaggeration — it should be read as agents simultaneously handling onboarding, coding, and monitoring, which in turn causes the volume of approvals and reviews landing back on humans to explode.

There's a lesson here for smaller teams and startups considering adopting agents: adding more agents doesn't mean you need fewer people — it creates a new kind of waiting time, where someone has to be available to make judgment calls whenever an agent gets stuck. Before following Grills's example of using agents instead of hiring, it's worth first measuring how many hours a day actually go into approvals and exception handling. Skip that step, and you risk ending up with one founder serving as the company's round-the-clock on-call system.

In the coming weeks, expect to see more signs of this kind of fatigue surface. It's likely we'll first see smartwatch apps for remotely controlling agents, or "agent manager" style products that handle overnight approvals on a founder's behalf. What's actually needed isn't a way to slow agents down — it's a way to safely defer certain decisions without requiring instant human intervention.