AI Brain Fry Is Real. And It's Coming for Your Best Employees.
Episode 20 explains how to adapt to the AI agent era without burnout from trying to multitask.
Last week, I had one of those days where I felt almost superhuman. I filed several state annual reports (our director of finance was in a cloud forest), wrote a bunch of marketing materials, pair programmed on some new features for GQueues, and in between, worked with Antigravity to build a prototype of an in-Gmail version of The Email Game that I then used to go through almost 200 emails that had piled up while I was at a conference. I had a good bit of help from others on our team, plus Claude and Gemini, but in addition to orchestrating the work, I did a lot of it myself. I felt great.
That night, though, I found myself red-faced, screaming at my six year old because he wouldn’t get ready for bed.
He had spent most of the night pushing boundaries, but that’s sort of what six year olds do. And normally, I have more patience before I hit my boiling point. I was really disappointed in myself, especially since I’d had such a great day at work.
It turns out that researchers have already given the condition I was experiencing a name: AI brain fry. Basically, AI brain fry means developing mental fatigue from excessive use and oversight of AI tools beyond sustainable cognitive capacity. After spending the entire day with my brain at max speed, switching back and forth between different projects, and using all the mini-breaks in the day to direct Antigravity, I had nothing left mentally. So when my son decided to spin his underwear above his head and then fling it at me instead of putting it on, I snapped.
This is now happening to a lot of people. (AI brain fry, obviously! I don’t have any data about boxer-brief projectiles.) A recent Harvard Business School survey of about 1,500 workers (a mixture of individual contributors and managers) found that heavy AI users are reporting exhaustion at levels they hadn’t seen before. The heaviest users are also disproportionately reporting that they want to quit their jobs. So even though the most ambitious AI adopters are likely getting more done than ever, they are also burning out faster than ever.
Our brains aren’t built for multi-tasking. So for the 2.5% of us who are “supertaskers” and can actually do multiple things at once, our new agentic world looks like paradise. The other 97.5% of us, well, we need some better ideas. And that’s what this episode is about.
Multitasking is stressful
Trying and failing to multitask is nothing new. Ever since computers became capable of running multiple programs at the same time, we’ve increasingly tried to do work on multiple things at once ourselves. And as the pandemic shifted our work communication to higher-interruption-level channels, the amount of time we worked on multiple projects at the same time increased too.
Unfortunately, true multitasking, where we think about two things at the same time, is an impossibility for 97.5% of us. The most informative research on this concept comes from the University of Utah, where researchers studied how well people could do math problems while driving. Weirdly, 2.5% of people actually got better at driving and better at math while doing them at the same time. The rest of the participants, though, did worse at both. The researchers theorized that instead of thinking about both processes at the same time, the vast majority of us have brains that rapidly switch contexts between one task and another each time our focus shifts. So what feels like “multitasking” is really “single-tasking with rapid switching.”
In 2008, Gloria Mark’s research group at UCI produced a study that found that rapid context switching very, very slightly increased productivity. But it came at a big cost: the participants who worked in an interrupted state reported a 2x increase in anxiety, a huge increase in stress, and a huge increase in the amount of effort that it took to do the tasks. Sounds about like my day last week!
That research group is also responsible for the often-cited statistic that it takes 23 minutes to make a full recovery from an interruption. That was true in their observation (participants would regularly take themselves down rabbit holes and work on other things before returning to where they started), but it doesn’t really give a full picture of how these context switches impact our attention.
More recent studies show that the cost of context switching scales with depth. The smallest possible interruption (a notification that is noticed but immediately ignored) costs approximately 7 seconds of attention. After that, the amount of attention lost scales linearly, taking 20-30 seconds to fully recover from a 5 second context switch, and about 5 minutes to fully recover from a 60 second one, based on the Memory for Goals model. Some fascinating research jointly produced by the US Air Force and makers of headphones with built-in EEG brain measurement equipment (I am now on the waiting list for this device!) confirmed the model’s accuracy in a real-world experiment.
For any context switch that takes over a minute, it takes around 5 minutes to fully reorient and regain full productivity on the previous task. And any deep disruption (anything over 5 minutes) basically takes you on the original 23-minute journey.
So with all of this lost time and increased stress, what if we just give up on multitasking and try to protect our focus? Between increased expectations around responsiveness and, even more importantly, the rise of AI agents, that’s no longer a winning strategy either.
AI agents and the fall of “just don’t multitask”
If you’ve been using AI to do hard things already, you’re familiar with the models having lag time between when you send them off and when they finish their work. Crowdsourced timing estimates for tasks like writing code include a 15-30 second “planning” period where AI is figuring out the context it needs, another 15-30 seconds of loading information into its context window, then several minutes to complete a quick-ish coding task. Deep research requests have a similar profile. Even the simplest requests require about a minute of waiting, and complicated requests take several minutes longer.
Nobody wants to just sit there and watch the thinking trace update – and it’s generally not a good use of time either. So we do something else and wait for a notification to come in to tell us that it’s done. Or, even more aggressively, maybe we kick off another AI agent to go work on something else for us at the same time.
All these context-switches add up to a lot of stress on their own. And the HBS study I referenced at the start of this post listed a few reasons beyond just the context switching that managing these AI tools are really hard on our brains.
First, AIs are not human, and it’s harder for humans to review AI output than to review human output. When humans do bad work, we tend to follow predictable patterns. For example, a lack of attention to detail will show up in the quality of the prose as well as in the quality of the reasoning. AIs don’t work the same way - the grammar is always flawless, and even things that don’t make any sense are written in ways that seem like they almost could. It takes a lot more cognitive effort to identify bad AI work product.
Second, the output from AIs is often a compressed version of all the inputs they took in. Because it takes an AI just seconds to “read” hundreds of pages and turn it into a few sentences of output, an AI-sourced research summary will include a staggering number of different sources given the amount of time it took to generate. A human doing the same project would have multiple days of reading time for their brains to process and sort and find connections between the different source materials. With AI assistance, you have to digest all of the summarized information all in one go.
Finally, because they can work on so many different things at once, working with AI can overload our working memory. Most people can only keep 5-7 facts or numbers or concepts in their mind at the same time. That’s fine if you have half an hour between context-switches or if the things you are working on are all related. But if you’re working on five totally different projects and redirecting your attention between them every minute as you juggle five simultaneous AI agents each awaiting your next instruction, your working memory will be overflowing dozens of times each day. And that’s exhausting.
The combination of more attention switches, across more diverse topics, with counterintuitive review processes, drawing from a huge diversity of sources all at once sounds maddening. Some of the heavy users of AI in the study reported feelings of paralysis, where they had so many possibilities and so many open threads that they just gave up and never shipped anything. Others adapted by “tokenmaxxing” where they tried to consume as much AI as humanly possible, using the parts of our brain that light up for games and competition to adapt to the agentic workflow. But whether they ended up paralyzed and accomplishing nothing, competing on tokens used rather than on quality or value of work produced, or just burned out, none of the most common adaptations seem good.
Given that multitasking, specifically cognitively-demanding multitasking, is going to be part of our work lives, how do we structure our days so that we have enough left in the tank to be good to our families and friends after the workday ends?
Subtask boundaries and open loops
There are three techniques that can all help us benefit from the productivity gains AI allows while also keeping us sane and helping our own contributions to projects stay high quality.
Use natural subtask boundaries
The single most valuable adaptation is to make AI wait for you, rather than letting it interrupt you. It turns out that our brains are really good at figuring out when the best times are to switch our focus. Research shows that, when we’re not externally interrupted, we switch our attention between tasks at subtask boundaries 94% of the time. Our brains automatically recognize “hey, I’m done with this part of this task; now would be a relatively cheap time to switch to working on something else.”
A study from Microsoft Research found that our recovery time is much quicker when we switch at these natural breakpoints. When workers got interrupted after finishing a sentence or a bullet point, they recovered 30% faster than when they got interrupted midway through. This study did not investigate what the recovery time would be when timing of the context-switch was fully user-driven, but it stands to reason that recovery times would be even shorter.
Applying a couple other techniques when you choose to switch your focus can help too. If you know you’re going to be coming back to your original task, keeping its window 75% or more visible while you work on the new task can help reduce both chain-of-diversion time and reloading where you were into your brain. And if you know it’s going to be a few minutes before you come back to the original task, leaving a “start here” breadcrumb, like the first few words of the next paragraph, or a brief note about what you were going to do next, can also make the shift in focus easier on your brain.
So when your AI system offers to notify you when it’s done, don’t let it! Instead, find a natural breakpoint in your own work, like when you’ve finished a paragraph of text or when you’ve finished a row of calculations in your spreadsheet. Then, when your brain is naturally ready to pause, you can go see what AI has ready for you. The AI doesn’t mind waiting - the hardware used for inference will be plenty busy helping someone else!
Limit open loops
If you’ve opened up any of the tools like Claude Code or Cowork that can run multiple AI projects at once, or even if you use more than one AI chatbot, it can be very tempting to look through your to-do list and decide that AI is going to work on everything on that list simultaneously, today. And it probably can make parallel progress on a whole lot of items on your list all at once!
The problem is that if you care at all about the deliverable, you’re going to also have to spend the day making parallel progress on all of those items. And for all the reasons we’ve already talked about, you’re not going to do a very good job at any of it, and you’re going to feel like crap at the end of the day.
Because our working memory can only hold 5-7 chunks and all of the open projects tax the same cognitive pool, trying to do everything everywhere all at once is doomed to fail.
Instead, keep in mind that your own energy and ability to think and understand will be the new limits on what you can do. And the Zeigarnik effect (stress increasing with the number of “incomplete” things your mind is keeping track of; the topic of our very first episode!) hasn’t gone anywhere. So instead of trying to do everything today, limit the number of projects you’re working on, and finish one before starting another.
Build routine-complex cycles
When you’re starting up a gym routine, you probably wouldn’t go with an every-day-is-leg-day approach. Your legs need time to recover from yesterday’s workout, so it’s better to make today work a different muscle group.
I don’t have good research to back this part up, but in my personal experience, our brains need a similar structure. When I’ve spent my morning doing really cognitively demanding tasks, especially if I’m doing more than one of them at once, I tend to have a much more productive afternoon if I spend it on routine or more creative projects, rather than jumping right back into something that is going to require a lot of squeezing on the exact same part of my brain. Working in cycles of 60-180 minutes of one “muscle group” and then rotating to a different one for the next period seems to help.
It’s not all AI’s fault
These strategies are built for the AI agent era, but even if you’re not using AI for anything, they still solve important problems in the ever-faster-paced world that we live in today.
Back in pre-social-media times, people used to spend an average of 2.5 minutes working in the same computer window before changing focus. Today, that’s down to 47 seconds. Without a doubt, there are more notifications and more input coming through chat apps than there were 15 years go, but even when there are no external interruptions, we still spend a lot less time working on one thing consecutively than we did before.
There’s evidence that our shortening attention spans aren’t just a behavioral change, too. Heavy usage of quick-attention-span apps can rewire your brain to have less gray matter. And taking away people’s phones slowly reversed the effect.
To some extent, we find ourselves in a reinforcement loop where our attention spans both working and at leisure are getting shorter, so we switch our focus even faster, which rewires our brains to have even shorter attention spans, and so on.
The answer can’t just be “don’t use AI.” Instead, it needs to be deliberate use of all of the tools that make multitasking so tempting.
Tip of the week
When we were preparing for this episode (and trying to figure out how to live in this world ourselves), we read dozens of articles about how people are adapting. AI read hundreds more for us.
One idea that stood out, attributed to a nameless marketing agency, was very simple: just be really dogmatic about taking a 5-10 minute break from multitasking every 90 minutes.
During that break, don’t use any screens or any AI tools, and definitely, definitely, don’t check any notifications. Walk around. Work on a pen-and-paper task. Read something in a book. Listen to something. Make a cup of tea.
Even after switching at subtask boundaries, limiting open loops, and alternating complex and routine tasks, it’s still easy to feel frazzled when working at an AI-dictated pace. As the philosopher Lao Tzu said, Nature does not hurry, yet everything is accomplished. Deliberately making time to slow down makes more sense than ever.
This post is based on Episode 20 of Less Busy Lab, a podcast of our field notes from 20 years running a productivity software company. Find this episode here, and subscribe wherever you get your podcasts!