Scattered Minds

A Morning With Eight Agents

It is mid-morning and eight agent sessions are open. One is refactoring a payment module, one is writing a migration, one is chasing a flaky test and the rest are somewhere in between. Every forty-five seconds or so you glance at one of them, read an intermediate result, decide whether it is on track and type the next instruction. Then you check mail until the next one wants you. Plenty of people skip the mail and start a ninth agent instead.

By lunch you have shipped more than on a normal day, and your head is full: it holds eight half-finished states, none of which you wrote and all of which you answer for, and there is no room left for a ninth thought.

Some people love this. It feels like a pleasant, slightly jittery flow, the closest thing to running a whole team from one chair. That reaction is real, and it stays real through everything below. The question is what this mode costs and where the limit sits.

What to Call It

We call it round-robin supervision. The agents work in parallel and you visit them one after another, the way a scheduler hands out time slices. Every visit lands on something half-finished, and your job there is to judge it. You have to understand work you didn’t do, and you answer for code you didn’t write.

The underlying problem is old. Human factors research has studied people supervising automated systems for decades and calls it supervisory control. Robotics even measured how many robots one person can run at once, and the answer depended on how autonomous the robots were and how hard the task was (numbers in the appendix). Agents aren’t robots, but the principle carries over: how many you can handle depends on how long each one runs without you. When an agent does need you, it usually wants a design decision, and that weighs more than a waypoint.

A study of how developers supervise coding agents found this demand in their online discussions of the practice: they have to keep enough understanding of work they did not write themselves (study in the appendix).

Why It Sets Some of Us Free

For me, this way of working is a release. For the first time I can work several lines of thought through properly, side by side. Results arrive fast. The current state of the art turns up by itself, and concepts get questioned and corrected along the way.

The mechanism is easy to name. I let go of the details, and the time it takes to test a creative idea in practice has collapsed. An idea gets tried almost as soon as I have it. That is what makes the work motivating for me.

This is not how everyone experiences it. The costs in the rest of this piece are real, and so is this.

What a Survey of AI Users Found

The most direct evidence so far comes from a Boston Consulting Group survey of US workers at large companies, published in Harvard Business Review in March 2026. The authors call the effect brain fry, mental fatigue from excessive use or oversight of AI tools, beyond what a person can cognitively handle, which participants described as a fog with trouble focusing and slower decisions. Engineering was among the more affected functions, and overseeing AI turned out to be the most taxing kind of use. People who reported brain fry also reported more decision fatigue and more frequent errors, and more of them intended to leave their job (all figures in the appendix).

The finding that matters most here concerns the number of AI tools used at the same time. Self-rated productivity rose from one tool to two, rose less with a third and dipped after three.

That curve compares different people at one moment, so it shows where the association turns and follows nobody who adds a fourth tool. Productivity and brain fry were both self-reported, and the lead authors work for a consultancy that sells AI transformation. In their data, AI use did not seem to predict increased burnout, which they keep apart from brain fry (caveats and chart).

Why Decisions Queue Up

Three findings from cognitive research explain why the curve turns.

The first is a bottleneck. In a brain-imaging study of people handling two overlapping simple tasks, early sensory processing ran largely in parallel, while the step where a response gets selected queued up and ran one task after the other (study in the appendix). The transfer to agents is ours: you can see eight dashboards at once and still decide about them one after another.

The second is residue. In experiments on attention residue, people struggled to pull their attention away from an unfinished task, and their performance on the next one suffered (study). An agent thread is unfinished every time you leave it.

The third is the clock. Information workers in one field study stayed on one screen for well under a minute. In a smaller, earlier study, people who came back to interrupted work the same day took on average more than twenty minutes to return to it, with other projects in between (numbers).

Set a polling cadence of thirty to sixty seconds, our own observation, next to that. It is far shorter than the time people were seen to need before returning to interrupted work, so a thread almost never gets picked up with a clean slate. We read this as a state that never settles, each glance adding residue before the last has cleared. We know of no study that has measured how long residue takes to fade, so that last step is our inference.

How Supervisors Lose the Picture

The slower cost takes longer to show. In 1983 Lisanne Bainbridge described the ironies of automation in industrial control rooms. Operators keep the tasks designers could not automate and take over in abnormal situations, exactly when skills that fade without practice are needed most. By taking away the easy parts of the job, automation can make the difficult parts harder (quotes in the appendix). The parallel to agents is ours.

Endsley and Kiris tested this in a 1995 experiment. Operators who no longer made the decisions understood the situation less well and took longer to decide once the automation failed, although their screens showed the same information. That is the agent operator’s position: you step in when the agent fails, the worst moment, holding the weakest picture of what is going on.

Tooling fixes part of it. In a Microsoft Research study, a view that tracks what each parallel coding agent is doing raised ticket throughput and made tracking feel less effortful, while the participants’ sense of control did not rise significantly (numbers). The study is a preprint by the tool’s own authors. Writing state down relieves working memory; the picture itself has to come from somewhere else.

Review follows the same slope. In a survey of knowledge workers, people with more confidence in the AI for a given task reported less critical thinking. That is a snapshot; a decline over weeks is our inference. Smelling a wrong intermediate step is a skill, and unused skills fade.

What the Research Does Not Show

We know of no study that has measured this working pattern with validated clinical instruments or followed people over months while they add agents. Nor do we know of a published capacity number for coding agents; the robotics fan-out figures are an analogy at best.

Individual differences are large. In a US Army simulation with unreliable automation, operators who rated their attentional control high suffered more from false alarms, and those who rated it low suffered more from missed detections. Which failure hurts depends on who is watching. In an interview study of developers with ADHD, one participant described concurrent agents helping them keep momentum, and others found manually auditing large volumes of AI code exhausting (both studies).

The strongest counter-position comes from a practitioner. Between March and June, Rémy Duthu of Mergify went from a cap of two agents to sometimes a dozen at once and felt less frazzled, because agent autonomy plus a status view turned babysitting into a queue he drained from the top. In the same post he concedes the core point: two interrupts at the same time cost what they always did, and his flood of pull requests made review more expensive for his teammates (his post). That moves the limit to a better place. What counts is how many threads need your judgement at the same moment.

Others will call this a moral panic, the complaint every new medium draws. The argument here rests on the serial bottleneck alone, which even Steven Pinker granted when he rebutted such panics (sources).

Where We Draw the Limit

What follows is our recommendation, drawn from the shape of the evidence. Nobody has measured it, and nobody yet knows exactly where the curve turns for coding agents.

Keep two to three threads that need real judgement. Four to eight are workable when most of them alert you on exceptions and only one or two sit in active review at a time.

Write state down: what each thread is doing, what it waits for, what done looks like. That relieves your head. The picture comes back through the last habit on this list.

Batch your reviews. Polling every forty-five seconds is a habit, and tooling can remove it. Look at agents when a batch is ready, on your schedule.

Leave the gap empty. The minute between two decisions is the closest thing to a pause your head gets, and mail fills it with one more unfinished thing.

And pull one thread all the way down to the implementation on a regular basis. Read the code, run it, change something by hand. Otherwise you turn into Bainbridge’s operator, the one who intervenes only when he knows least.

Why We Track This

We run this pattern every day, in client engagements and in our own work. We have hit the wall ourselves: the afternoon where every glance costs more than the one before and the reviews quietly get softer. We needed a name for that state before we could set a limit on it.

Appendix: Numbers and Sources

Supervisory control and fan-out

• Supervisory control. Human factors research calls the supervision of automated systems supervisory control, a field Thomas Sheridan did much to shape.
• Robot fan-out. Fan-out is the number of robots one person can effectively run at once (Olsen & Goodrich, 2003). With simulated robots it ranged from about one to nine, depending on how autonomous the robots were and how hard the task was (Olsen & Wood, CHI 2004).
• Supervising coding agents. Park et al. watched 19 experienced developers work with coding agents, built a model of supervision from what they saw and applied it to 102 Reddit threads in which developers discuss their practice. In those threads they found “a supervisory demand to maintain enough understanding of work they had not written themselves” (arXiv:2609.24234). The ParallelPilot study from Microsoft Research (see Automation and situation awareness) also frames parallel AI coding as supervisory control.

The BCG survey on brain fry

Bedard et al., “When Using AI Leads to ‘Brain Fry’”, Harvard Business Review, March 2026: a Boston Consulting Group survey of 1,488 full-time US-based workers at large companies.

• Definition. The authors call the effect “AI brain fry”: “mental fatigue from excessive use or oversight of AI tools beyond one’s cognitive capacity.” Participants described “a ‘buzzing’ feeling or a mental fog with difficulty focusing, slower decision-making, and headaches.”
• Prevalence. Of the respondents who use AI at work, 14% reported it, from 6% in legal roles to 26% in marketing, with engineering among the more affected functions.
• Oversight. Oversight was the most taxing kind of AI use: high oversight went with 14% more mental effort, 12% more mental fatigue and 19% more information overload than low oversight.
• Decisions, errors, quitting. People reporting brain fry also reported 33% more decision fatigue and scored 11% and 39% higher on minor and major error frequency. Of them, 34% intended to leave their job, against 25% of the rest.
• Number of tools. Self-rated productivity rose from one AI tool used at the same time to two, rose less with a third, and then: “After three tools, though, productivity scores dipped.”

• Caveats. A cross-sectional survey compares different people at one moment: it shows association only and follows nobody who adds a fourth tool. Productivity is self-reported, brain fry was one question about whether people had ever felt it, and the lead authors work for a consultancy that sells AI transformation. We have found no longitudinal or experimental study of this exact pattern.
• Burnout. The authors keep brain fry apart from burnout: in their data, AI use “didn’t seem to predict increased burnout.”
• A line they quote. One line they quote fits our reader closely. Francesco Bonacci, posting on X, described ending each day exhausted, “not from the work itself, but from the managing of the work.”

The cognitive bottleneck

• Serial response selection. Yue, Newton and Marois used ultrafast 7-Tesla fMRI on 26 young adults handling two overlapping simple tasks (Nature Communications, 2025). Early sensory processing unfolded “largely in parallel.” The task-specific work in the fronto-parietal multiple-demand network, where a response gets selected, queued up and ran serially. The transfer to agents is ours.
• A practitioner’s version. Addy Osmani said it plainly in April 2026: “Your cognitive bandwidth doesn’t parallelize.”
• Attention residue. Sophie Leroy showed in two experiments that people struggle to pull their attention away from an unfinished task, and that their performance on the next one suffers (OBHDP, 2009). Only finishing the first task under time pressure helped people let go. Leroy did not measure how long residue takes to fade, and we know of no study that has.
• Time on one screen. Gloria Mark and colleagues found information workers staying on one screen for a median of 40 seconds (CHI 2016).
• Returning to interrupted work. In an earlier study, people who came back to interrupted work the same day took on average 25 minutes 26 seconds to return to it, with 2.26 other projects in between (Mark, Gonzalez & Harris, CHI 2005; 24 people at one company, wide spread).
• Polling cadence. The cadence of thirty to sixty seconds is our own observation, and the step from these findings to a state that never settles is our inference.

Automation and situation awareness

• Ironies of automation. In 1983 Lisanne Bainbridge wrote “Ironies of Automation” about industrial control rooms (Automatica). Automation, she argued, “may expand rather than eliminate problems with the human operator.” Operators keep the tasks designers could not automate and take over in abnormal situations, exactly when skills that fade without practice are needed most. “By taking away the easy parts of his task, automation can make the difficult parts of the human operator’s task more difficult.” The parallel to agents is ours.
• Out of the loop. Endsley and Kiris tested this in 1995 (Human Factors). They automated a navigation task at five levels of control. Operators who no longer made the decisions understood the situation less well and took longer to decide once the automation failed, although their screens showed the same information. The authors trace this to the shift from active to passive processing.
• ParallelPilot. In the ParallelPilot study from Microsoft Research, 16 participants used a coordination and monitoring view for parallel coding agents (Long et al., arXiv:2609.33113). Ticket throughput rose 63%, from 0.272 to 0.445 tickets per minute, and peak concurrent agents went up by about one (2.31 to 3.25). Self-rated tracking effort fell from 4.63 to 2.75 on a seven-point scale. Perceived control and perceived intervention success did not rise significantly; the authors report that the bookkeeping the tool removed had “reinforced their sense of control.” It is a preprint by the prototype’s own authors, tested on experienced developers who were mostly new to running agents in parallel.
• Confidence and critical thinking. In a survey of 319 knowledge workers, higher confidence in the AI for a given task went with less self-reported critical thinking, and higher confidence in one’s own ability went with more (Lee et al., CHI 2025). That is a snapshot; a decline over weeks is our inference.

Individual differences and counter-positions

• Attentional control. In a US Army simulation with unreliable automation, operators who rated their attentional control high suffered more from false alarms, and those who rated it low suffered more from missed detections (Chen & Terrence, Ergonomics, 2009).
• Developers with ADHD. In a 2026 interview study of 14 developers with ADHD (formally diagnosed or self-identified), one participant described using concurrent AI agents to maintain momentum. Participants also described manually auditing large volumes of AI code in sensitive or high-stakes production settings as increasing cognitive exhaustion and anxiety over hidden bugs (Pimenova et al., arXiv:2609.21254).
• A practitioner’s counter-position. Rémy Duthu of Mergify capped himself at two agents in March and by June ran “sometimes a dozen at once,” feeling “less frazzled,” because autonomy plus a status view “turn babysitting into a queue I drain top-down.” He concedes the core in the same post: “Two simultaneous interrupts cost exactly what they always did,” and code review got worse: his flood of PRs made review more expensive for his teammates.
• Moral panic. In 2020 Amy Orben traced the recurring cycle of panics that every new medium draws. Steven Pinker, in his 2010 rebuttal of such panics, granted the serial bottleneck: “Genuine multitasking, too, has been exposed as a myth.”