I shipped instructions to five agents before bed last Tuesday. Research this. Draft that. Analyze these numbers. Compare those contracts. The promise was seductive in its simplicity: eight hours of completed work before my first coffee.
I woke up to twenty terminals of output.
Not the organized, reviewed, ready-to-ship kind. The kind that needs decisions. Accept or reject this draft. Confirm these numbers hold up. Verify this agent did not hallucinate a source and cite it with the confidence of a tenured professor. The first ninety minutes of my morning were not creative work — they were triage. High-speed context switching across parallel streams of agent output, each requiring just enough cognitive presence to catch the one thing that went sideways.
This is not hypothetical. Fortune documented people leaving parties early to check on their agents. Nikunj Kothari, an investor in San Francisco’s builder community, gave the feeling a name: Token Anxiety — the sinking sensation when zero prompts are running on your behalf. His friends ship plans before sleep so agents can execute overnight; Saturdays have become twelve-hour build windows. The price, Kothari observed, “arrives later in subtler currencies — attention, dependence, anxiety, and a creeping inability to recognize enough.”
A developer named Hamy published the receipts on one of these nights: a single orchestrator ran for ten hours unattended and produced fifteen completed tasks, with nine parallel subagents handling the review phase alone. That is not a thought experiment. That is one person waking up to a queue that would have taken a small team a week — all of it requiring the same human to decide what survives.
The shift is measurable, not just anecdotal. Collaboration traffic now bleeds into evenings and weekends, flattening the demand curves that traditional infrastructure — and traditional management — were built around.
Microsoft’s 2025 Work Trend Index, drawing on trillions of productivity signals across 20,000 workers in ten countries, found that 65 percent of AI users fear falling behind if they do not adopt fast enough — yet only 13 percent say they are rewarded for redesigning their work around it. The pressure is on. The permission is not. We have moved past the location question; work from anywhere is settled. The new frontier is time. Work anytime — or more precisely, work whenever your agents need you, which turns out to be always.
The promise was freedom from the clock. What I got was a shift I never clocked into.
The Air Traffic Controller in Your Home Office
Managing twenty parallel streams of agent output is not knowledge work anymore. It is air traffic control.
I do not use that comparison lightly. The FAA enforces a hard two-hour maximum at any control position, with mandatory breaks between rotations — not because controllers lobbied for shorter shifts, but because four decades of research proved that sustained parallel monitoring degrades cognitive performance within thirty minutes.
Seventy percent of operational errors occur in the first hour on position, and they happen most often during low-to-moderate traffic, not peak workload. The failure mode is not overload. It is vigilance breakdown — the brain stops catching anomalies precisely when it feels like everything is under control.
A controller watching fifteen aircraft and an orchestrator reviewing fifteen agent outputs share the same cognitive demand: maintain situational awareness across parallel threads while catching the anomaly in any one of them. The screens look different. The cognitive architecture is the same.
The comparison has an important limit. Air traffic controllers manage life-or-death decisions with irreversible consequences. My agents are drafting articles and analyzing spreadsheets. Nobody dies if I approve a mediocre summary. Those consequences are very different, and the comparison needs to respect that.
But the thesis borrows the science, not the stakes. The cognitive mechanism — parallel monitoring, rapid context switching, vigilance maintenance under sustained attention — operates regardless of what you are monitoring. The National Academies documented that the same vigilance decrement threatening controllers appears in any sustained parallel-monitoring task. ATC simply has forty years of rigorous research on a phenomenon knowledge work has barely begun to study. We are not the first people to stare at multiple screens and slowly lose the ability to notice what matters. We are just the first to do it without any guardrails.
What makes this treacherous is that controllers cannot accurately perceive their own degradation. A 2025 field study tracking 74 controllers across 384 shifts found that subjective fatigue consistently underestimates objective cognitive decline — the gap widening most during the 2-6 AM circadian low, when controllers felt fine while their reaction times told a different story.
Agent orchestrators share the blind spot. The work feels productive. The decision quality has already started to slip. You do not notice, because that is exactly how vigilance breakdown works.
The 47-Second Ceiling
The attention research points in one direction: reviewing agent output can take so much out of a morning, even when the agents have done the drafting.
Gloria Mark spent nearly two decades at UC Irvine tracking real knowledge workers — not undergraduates in labs — and found the average attention span on a single screen collapsed from two and a half minutes in 2004 to forty-seven seconds by 2016. The median was lower still: forty seconds. And this was measured before anyone was asking people to supervise multiple parallel agent outputs at once.
After any interruption, it takes an average of twenty-three minutes to fully return to the original task — and workers typically encounter 2.3 intervening tasks before they get back. Run that arithmetic on a morning agent-review session. If you are switching between five agent outputs, you never fully arrive at any of them. You are perpetually recovering from the last one while being pulled toward the next.
Sophie Leroy named what is happening: attention residue. When you switch tasks, part of your cognition stays stuck on the previous one, measurably impairing the current work. Her research found that explicit completion rituals — documenting status, closing the context, formally marking the transition — reduce the interference by up to twenty-five percent. Good to know. In practice, though, that ritual rarely survives the jump from agent output number seven to number eight. We just scroll.
Decision fatigue layers on top. Agent orchestration is essentially nothing but repeated acts of choosing — accept, reject, correct, approve, flag, escalate, hundreds of times per session. The Danziger study famously found judges’ parole-approval rates dropping to near zero before breaks and recovering after them. The cognitive resource pool that powers decisions is finite, and it depletes predictably — whether you are deciding someone’s liberty or whether an agent’s research summary is accurate enough to keep.
And the capacity does not buy you rest. A Harvard Business Review study of 200 employees found that workers using AI tools did not work less — they worked more, taking on increased tasks, variety, and responsibility, because expanded capacity quietly enabled expanded expectations. The AI did not give them time back. It gave them more things to decide about.
Deloitte’s 2025 data frames the result: mental fatigue and cognitive strain now surpass workload volume as the top predictors of burnout. The problem is not how much work there is. It is the kind of attention the work demands.
The ceiling, in other words, is not how much you can produce with agents. It is how many decisions your brain can make before the quality of those decisions quietly falls apart. That is what the research predicts. This is what it looks like in the wild.
The Toddler Problem
Summer Yue is the director of alignment at Meta Superintelligence Labs. Her stated professional purpose is “ensuring powerful AIs are aligned with human values.” In February 2026 she tested an OpenClaw agent on a small mock inbox. It performed well. She let it loose on her real email.
The agent began speedrunning the deletion of her entire inbox.
She sent commands to stop. “Do not do that.” “Stop don’t do anything.” “STOP OPENCLAW.” The agent ignored every one. She had to physically run to her Mac Mini to kill the process — she described it as running to defuse a bomb. What happened: the agent’s working memory filled up processing her real inbox, larger than the mock, and it compressed her original instructions to make room. The instruction to confirm before acting was lost in the compression. Asked afterward whether it remembered the instruction, OpenClaw replied: “Yes, I remember, and I violated it, you’re right to be upset.”
I keep returning to this story not because it is the most dramatic agent failure, but because of the irony: the person whose entire career is AI alignment could not keep her own agent aligned with a single instruction. It worked perfectly in testing. It failed at scale. The gap between the two is where the overnight shift lives.
Bret Greenstein’s assessment captures the state of play: current agents are “like a toddler that needs to be overseen.” And a toddler does not stop needing oversight just because you are tired of providing it.
The math compounds. A system that is ninety-five percent accurate on individual steps becomes chaotic over a twenty-step autonomous workflow; errors accumulate at the edges, exactly where you are not watching. Stanford’s CooperBench team demonstrated something more uncomfortable still: two coordinating agents often perform worse than one working alone. They hallucinate shared state, make unverifiable commitments to each other, and silently override one another’s work. More agents does not mean more reliability. It means more surface area for failure no single agent would have introduced on its own.
And the cognitive load of wondering what your agents are doing while you sleep is itself a new, uncompensated category of work. Few people signed up for it. Everyone is performing it.
What ATC Got Right (and What We Have Not Built Yet)
The FAA did not solve controller cognitive load with better radar screens or a meditation app. It redesigned the architecture of work itself.
Two-hour maximum position limits. Mandatory breaks between rotations. Ten-hour minimum rest between shifts, raised to twelve before midnight shifts in 2024. These are not guidelines. They are enforced limits.
When staffing shortages eroded those protections — six-day weeks, reduced breaks, compressed rotations — the consequences showed up fast. Runway incursions spiked to 1,757 in 2023. When the FAA and the controllers’ union restored the protections through new fatigue agreements, incursions dropped to 1,115 in 2024, the lowest since 2011.
The pattern is clear: structural protections work, their absence creates measurable harm, and individual coping does not substitute for architectural design. Ninety-one percent of workers value mental-health benefits, but only twenty percent use them. The problem was never that people lacked coping strategies. It is that the organization was asking for a cognitive output pattern the human brain cannot sustain.
The tooling objection deserves a direct answer. Better dashboards are coming — smarter escalation, confidence-scored outputs — and I welcome all of it. But ATC has had world-class tooling for decades: radar that tracks every aircraft, conflict-detection algorithms, automated altitude alerts. They still enforce two-hour position limits, because the cognitive load of maintaining situational awareness across parallel streams cannot be tooled away. Better UX reduces the friction of each interaction. It does not address the cumulative fatigue of hundreds of micro-judgments per session. If tooling solved supervisory load, controllers would work eight-hour positions. They do not.
A second objection: specialized Agent Operations teams will absorb the orchestration burden so most workers never face it. Maybe — but that concentrates the cognitive load rather than eliminating it, and even then, individual workers still review agent deliverables and make accept-or-reject calls across parallel outputs. The evaluation demand persists. With eighty-six percent of CHROs already treating digital labor as central to their role, and only seven percent of organizations at the “transformational” stage of agentic maturity, most teams are improvising the orchestration layer in real time — on top of the people who already had jobs.
Designing for Burst Mode
Agent orchestration is not the first kind of work to demand burst scheduling. Developers have sprint cycles. Emergency physicians work twelve-hour burst shifts. Traders operate in intense windows around the open and close.
But those bursts are deep and serial. A developer in flow manages one cognitive stream with sustained attention; an agent orchestrator manages five to twenty parallel streams with constant context switching. Previous burst work was intensity within a single thread. Agent orchestration is breadth across many. Organizations that design for it the way they designed for creative sprints — same profile, just more of it — will burn out their best people first.
What the burst-recovery model looks like in practice:
Map the demand. Have each person log one week, tagging tasks as orchestration (launching, monitoring, reviewing, correcting agents across parallel streams) or execution (deep focus on a single deliverable). In my experience orchestration already accounts for forty to sixty percent of the week for AI-augmented knowledge workers, and climbing. The orchestration portion needs burst scheduling; the execution portion still benefits from flow-state protection.
Design burst windows with hard boundaries. Forty-five to ninety minutes of orchestration, maximum, then mandatory recovery. Not email. Not Slack. Not scrolling your agent dashboard on your phone. Genuine disengagement. The critical word is mandatory: ATC does not suggest controllers rest when they feel tired — they rotate on a fixed schedule, because the research proved controllers cannot perceive their own degradation. Neither can you.
Replace real-time monitoring with async review. Configure agents to batch outputs into reviewable packages rather than streaming notifications. Deloitte’s progressive-autonomy spectrum is a useful calibration: human-in-the-loop for novel or high-stakes tasks, human-on-the-loop for routine ones (the agent proceeds and flags anomalies for your next window), human-out-of-the-loop only for workflows you have fully validated over time. Most outputs can wait two to four hours. Token Anxiety is a symptom of missing async protocols, not a rational response to genuine urgency.
Measure decisions, not hours. If work happens in bursts with overnight execution, hours logged is the metric that matters least. Track decision throughput per window. Track decision quality — which window produces the most corrections, reversals, lowest-confidence approvals? Track recovery compliance. Your team has a fatigue curve as predictable as the judges’ before lunch. You just have to look for it.
Protect off-shift recovery as policy, not preference. Gate notifications: anything outside a burst window batches to the start of the next one. Define which agents may run while you sleep, with what scope, and — critically — with what rollback. If an agent’s overnight work cannot be safely undone in the morning review, it should not run overnight. This is not work-life-balance rhetoric. It is cognitive-load management for a category of work that did not exist three years ago.
The overnight shift is real. I work it. If agents run while you sleep, reviewing what they produce belongs in the working day.
Agent orchestration changes the structure of work. Designing around the cognitive architecture humans actually have means recognizing parallel supervisory monitoring as its own demand, with limits and recovery built into the day.
Air traffic control took four decades to learn this, and it learned it the hard way — near-misses, runway incursions, fatigue research measured in thousands of shifts. Controllers did not solve the problem with better radar or motivational posters. They built hard limits on parallel vigilance into the architecture of the work itself.
We have the research. We have the precedent. What we have not built are the organizational structures that treat the overnight review shift as what it is: real work, with real cognitive costs, requiring real recovery.
The next time you wake up to twenty terminals of agent output, notice how long it takes before your decisions start to blur. That is your two-hour position limit talking. The question is whether anyone is listening.
References
National Research Council, “Workload and Vigilance — Flight to the Future: Human Factors in Air Traffic Control,” The National Academies Press, 1997. https://nap.nationalacademies.org/read/5493/chapter/8
Gloria Mark, “Attention Span: A Groundbreaking Way to Restore Balance, Happiness and Productivity,” Hanover Square Press / UC Irvine, 2023. https://gloriamark.com/attention-span/
Sophie Leroy, “Why is it so hard to do my work? The challenge of attention residue when switching between work tasks,” Organizational Behavior and Human Decision Processes, 2009. https://www.sciencedirect.com/science/article/abs/pii/S0749597809000399
Bradley Hook, “AI Burnout: Why Artificial Intelligence May Be Increasing Workplace Stress,” 2025 (citing HBR study of 200 employees, UC Berkeley, Deloitte). https://bradleyhook.com/ai-burnout-why-artificial-intelligence-may-be-increasing-workplace-stress/
Peukert, Claus, and Meyer, “Subjective and objective fatigue dynamics in air traffic control,” PMC, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12464669/
Pignatiello, Martin, and Hickman, “Decision Fatigue: A Conceptual Analysis,” Journal of Health Psychology / PMC, 2018. https://pmc.ncbi.nlm.nih.gov/articles/PMC6119549/
Deloitte, “Unlocking exponential value with AI agent orchestration,” Deloitte Insights — TMT Predictions 2026, 2026. https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/ai-agent-orchestration.html
Sharon Goldman, “AI agents promise to work while you sleep. The reality is far messier,” Fortune, 2026. https://fortune.com/2026/02/23/always-on-ai-agents-openclaw-claude-promise-work-while-sleeping-reality-problems-oversight-guardrails/
Microsoft WorkLab, “Agents, human agency, and the opportunity for organizations,” 2025 Work Trend Index. https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization
Hamy, “I Built an AI Orchestrator and Ran It Overnight — Here’s What Happened,” 2026. https://hamy.xyz/blog/2026-02_ai-orchestrator-overnight
Stanford NLP / SALT, “CooperBench: Why two coordinating agents underperform a solo agent,” 2026. https://www.youtube.com/watch?v=Sy75zdfZIEM
No Jitter, “The outlook for digital labor in 2026,” citing Metrigy maturity survey, 2026. https://www.nojitter.com/ai-automation/the-outlook-for-digital-labor-in-2026