AI Agents Push Startup Founders Into Longer Hours, Not Shorter

The productivity tool that never clocks out

AI agents were sold as a way to take work off people's plates. For a group of startup founders, they have done the opposite. A Wall Street Journal feature by Katherine Bindley, picked up widely across tech media in late August, describes founders who manage fleets of generative-AI agents and now work longer and more erratic hours than before. Their reason is simple: they don't want those agents sitting idle.13

The article opens on Aditya Sharma, co-founder of Keel, which makes AI troubleshooting software for IT teams. He recently stayed up until 6 a.m. working with his agents, which build customer features, do research for an in-house model and monitor marketing campaigns.1 Sharma's reasoning sums up the problem. An agent can finish its task at any hour, so leaving it blocked overnight for eight hours feels too costly.1 He also says he has gained weight in recent months and finds it harder to be present with people outside work, because the agents keep pulling at his attention.1

The other founders tell similar stories. Peter Pezaris had promised his wife he would retire after four successful exits. Instead he launched Proxon, which makes AI workforce-management software, about four months before the story ran. He typically works from 7:30 a.m. to 2 a.m.1 He estimates his six-developer company moves 30 times faster with agents than it would without them, and he compares working with them to a drug.1 Ajay Kalia, a former member of Spotify's AI group who now runs a startup called Alt, works roughly from 5:30 or 6 a.m. until 9 or 10 p.m. He approves agent requests from his Apple Watch, which buzzes about every ten minutes, including when he is out running.1

Why more output creates more work

The reporting shows how this happens. Agents do not just finish work. They create more of it for the humans around them. In Pezaris's case, signing up customers faster means more customer requests that people have to answer.1 One newsletter that summarized the piece called it a productivity loop with no natural stopping point: each new capability opens another task, another product or another customer to support.2

The second driver is psychological. Pezaris describes a fear of missing out that compounds: any minute he isn't working feels like a lost week of agent output.1 Kalia says new models let him take on far more ambitious projects, but he is never done, and every few months he has to rethink his work around what the agents can now do.1 Abby Grills, co-founder of the data-gathering platform Riveter, shows a quieter version of the same pattern. She and her co-founder are the company's only two employees. They keep planning to hire, then find an agent can handle the next task instead.1

One point in the original reporting is easy to miss. Not every founder has given in to the extremes. Grills works about ten hours a day, takes one weekend day off, and says she and her co-founder have become more careful about resting, even though she is tired all the time.1 She saw the cost of overwork during Y Combinator's 2024 program, when a founder in her cohort was hospitalized.1

Where the coverage diverges

Secondary outlets told the story in noticeably different ways. Techmeme's summary called the phenomenon "productivity FOMO" and noted that commentators found the irony appealing: tools meant to save time are stretching the workday.3 One aggregator went further, saying founders are working 90-hour weeks managing agent fleets. That figure is a stronger, more specific claim than the founder schedules in the original article.6 The same aggregator connected the story to Berkeley Haas research from February and to separate WSJ reporting, both finding that workers' estimates of how much time AI saves them differ sharply from what executives claim.6 The Scale and Strategy newsletter took a more essay-like view. It argued that when work can continue indefinitely, the limit on a company is no longer what it can produce but whether its people can make themselves stop.2

The best way to read the story is not as a burnout scare piece. It is early evidence of a structural problem: autonomous agents move the bottleneck in a company from execution to human supervision. That reading is supported by earlier reporting from Europe. In March, Sifted described founders who felt stuck in a "permanent product launch cycle" as they adopted tools like Claude Code and OpenClaw.7 Phily Hayes, CEO of the agent developer Zerve, told the outlet that AI has made individuals faster without making companies calmer, and that teams are attempting more rather than working less.7 Ted Chalouhi of Duku AI said his nine-person team can do the work of 25 but that he had not been on a date in months. Lean Health co-founder Jasmine Sayyari said she is constantly burnt out and afraid of falling behind.7 In other words, the WSJ story describes a pattern that has been building for at least half a year, not a one-off.

The enterprise mirror: oversight is the real constraint

What founders are going through personally, large enterprises are going through at the level of governance, which makes the founder story useful for the wider autonomous-agent market. Headline adoption numbers look strong. KPMG's Q3 2026 survey found 62% of large US businesses building, developing or deploying agents, and 25% working on multi-agent systems.11 Gartner has forecast that task-specific agents will appear in 40% of enterprise applications by the end of 2026, up from under 5% in 2025.14

The production numbers are much weaker, and the research firms disagree about them. One compilation notes that while McKinsey finds 62% of organizations at least experimenting with agents, a separate Deloitte count puts only 11% in active production.20 A ServiceNow index found that just 9% of enterprises are making meaningful progress on autonomous multistep workflows, even though 59% say they use agentic AI.19 A Contentstack survey of 621 digital leaders was more optimistic, with 58% reporting programs in production.15 Ramp transaction data suggests only 16% of enterprise AI deployments qualify as true agents. The rest are fixed-sequence or routing workflows.13

The spread comes partly from definitions, but the common thread is that the hard part is supervision, not technology. A Codiste survey of CTOs put the median technical build time for an enterprise agent at 4.2 months, compared with 7.2 months for governance and compliance.18 A Forrester study commissioned by Boomi found that only 34% of organizations trust the actions their agents take, and Deloitte estimates that only about 20% have a mature governance model for autonomous agents.18 Gartner also predicts that more than 40% of agentic projects will be canceled by the end of 2027.11

Put simply, enterprises are bringing in compliance teams, audit logs and approval checkpoints to do the job that Kalia does from his wrist on a run. Industry guides now recommend permission boundaries, human checkpoints for high-risk actions and new roles such as agent orchestrators to manage handoffs between agents and people.16 Founder-focused guides make the same point from the startup side. Agents need ongoing oversight because their logic can drift away from company goals, so using them changes management work rather than removing it.8 One guide says the most underrated part of an agent stack is monitoring and evaluation, and that this is where most engineering time goes in a one-person company.10

The economics behind the grind

The financial incentives explain why founders keep pushing anyway. Promoters of agent-native startups say a full agentic stack can cost a few hundred dollars a month while replacing functions that once took tens of thousands of dollars in annual payroll, though that is a marketing claim, not audited data.4 Fortune's reporting on solo founders offers a more cautious view. Monthly AI bills at lean startups running always-on agents can reach hundreds of thousands of dollars, sometimes comparable to the salaries they replace. Compute does scale more flexibly than staff, though, and founders keep more equity.5 An NYU Stern professor quoted by Fortune warned that agents cannot replace specialist judgment, leaving founders to take the output partly on faith.5 One founder even shut down his customer-support bot because answering tickets himself kept him close to the product.5

This tension is what drives the WSJ story. Small teams can now produce what used to take a department, which keeps the founder involved in every decision the agents cannot make alone.

What to watch

The lesson for the enterprise market is that adding more autonomy does not reduce the need for human attention. It concentrates that attention on a few overloaded people. The organizations that do well are likely to be the ones that clearly define which decisions agents may make on their own, which ones go to a human, and when. One founder guide calls that line harder to build than any workflow.10 Until that line is clear, the human in the loop will remain the limiting factor, whether it is a compliance committee at a bank or a founder in San Francisco checking on agents at 3 a.m.