Your AI Agents Never Sleep. The Approval Queue Still Eats Your Nights.
Agents can work around the clock, but every checkpoint where they wait for a human becomes a bottleneck that runs on your schedule. What the reporting inside agent-first startups found, and the design patterns that put the queue back in its place.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
The pitch for AI agents is easy to repeat: they do not sleep, so work continues while you do. Founders inside agent-first startups are living in the second half of that sentence, and the view is worse than the brochure. The agents stay up. So does everyone supervising them.
The Wall Street Journal talked to founders running companies on top of AI agents, and the Spanish tech outlet Xataka summarized the pattern in a piece titled, roughly, "It's like a drug: AI agents never sleep, so startup founders don't either." The details are worth sitting with, because they describe a failure mode that will reach anyone who wires agents into real work, not just venture-funded teams in San Francisco.
What the reporting actually found
Aditya Sharma, cofounder of a startup called Keel, described days that end at six in the morning because he was up handling his agents' requests. Another founder, Ajay Kalia of Alt, puts in fifteen-hour days and still gets agent permission requests pushed to his Apple Watch mid-run. He told the Journal he is not sure it is healthy to approve an agent's action while checking his watch during a jog. Peter Pezaris of Proxon put it more bluntly: it is like a drug, he estimates his company runs roughly thirty times faster with agents, and his working day now stretches from 7:30 in the morning to two at night.
None of this is mysterious once you look at where the hours go.
Agents block on humans, and blocking is expensive
An agent can execute a long workflow, but most deployments are not fully autonomous. Certain steps require a human yes: spending money, sending something to a customer, writing to a production system, deleting data. When the agent reaches one of those steps, it stops and waits. If nobody answers, it sits there.
Sharma's framing, as reported, is that the cost of agents sitting blocked for eight hours is too high. That single sentence explains the broken sleep. An idle agent is not just paused; it is a row of blocked downstream tasks, a stalled batch, a quota sitting unused while the clock runs. So the human supervisor stays reachable, always, just in case the queue lights up.
Layered on top is plain competitive fear. Pezaris again: every minute he is not working, he loses the equivalent of a week of work. When you believe your agent fleet multiplies your output thirtyfold, sleep starts to look like an expensive hobby. Kalia says everyone in his network quietly admits the pace is unsustainable, and then keeps it up anyway, because stopping while competitors do not feels worse.
The uncomfortable summary: the human-in-the-loop pattern was supposed to keep humans in charge. In practice, the machine batch-processes work and the human becomes a single-threaded, always-on interrupt handler.
The fix is structural, not motivational
Advice like "set boundaries" does not survive contact with a blocked queue at 3 a.m. The founders in these stories know the pace is unhealthy. The problem is that their systems make every hour of delay expensive. Change the system instead.
Tier your gates by blast radius. Not every agent action deserves a human yes. Reading, drafting, searching, and generating candidates can run default-open only for allowlisted, non-sensitive resources, with a log. Spending money, publishing externally, sending messages, and anything irreversible run default-blocked. Most agent setups that destroy sleep have one undifferentiated approval mode for everything, which means humans are paged for decisions a rule could make.
Batch the queue instead of streaming it. An agent that pings your watch per request trains you like a slot machine. An agent that accumulates approvals into a digest, reviewable in two sessions a day, returns your attention to you at the cost of latency: throughput holds roughly steady for independent, non-time-sensitive approvals, while dependent or deadline-bound work can still stall. If your tooling cannot batch, that is a tooling decision to revisit, not a personal discipline problem.
Let agents run dry overnight. The "eight hours blocked is too expensive" calculation deserves a second look. Often the expensive part is not the idle time but the context rot and the retry chaos of half-supervised overnight runs. Give the agent a bounded nightly plan, let it exhaust the work that needs no approval, and review the blocked tail with coffee. The marginal cost of one night of queue latency is usually far below the marginal cost of a founder who never sleeps.
Measure utilization of the queue, not of yourself. Track how long approvals wait, how many were pure rubber stamps, and what fraction of blocked actions were low-risk. A high rubber-stamp rate means the gate is in the wrong place. Move it.
The creative-work version of the same trap
If you build images or video with AI rather than run agents, you already know this queue. A generation pipeline produces twenty candidates and every one of them waits for your eye. The model works in seconds; the selection and judgment stack up into the hours you actually feel at the end of the day. The same structural fix applies: separate the pass that a machine or a preset can judge from the pass that genuinely needs your taste, and schedule the second pass instead of letting it interrupt the first one.
The agents will keep working at night either way. The question is whether your approval is genuinely on their critical path, or whether you just never got around to writing the rules that would let them proceed without you. Most of the time, it is the second one.