AI Brain Fry Is a Design Problem

Agent management creates cognitive strain. Even slightly better design could help.

If you’re working with AI agents, you know the feeling. It’s like your brain is biking through mud.

You get Claude working on a task, and you know it’ll take a few seconds. While you’re looking at thought traces, an idea for a related task pops into your head. You open a new chat, and start Claude working on the second task.

If you’re like me and Sarah, this multiplies. You might have 4-6 concurrent tasks running at the same time. Some of them related, some not. You’re switching between the chats, and every time you do, you need to reorient yourself to what you’re looking at. You have to read paragraphs of text, check output files, identify errors and fix them. The outputs are 80% good, but fixing that remaining 20% takes a lot more effort than you'd expect.

And while all of this is happening, you’re using an interface that was not designed for this behavior. It’s designed for linear, one-at-time, short-chat conversations.

As a result, you start to get a bit disoriented. You return to your first chat and think — wait, what was this about?

Humans are notoriously bad at multitasking. With agentic AI, the cognitive impacts are even more severe. You’re left feeling scattered, foggy, and exhausted.

This is your brain on agentic AI. (Photo by Saurabh Solanke)

Claude’s Frying Our Brains

This is AI brain fry.

In March 2026, Boston Consulting Group published a study of 1,488 U.S. workers that identified a distinct pattern of mental fatigue tied specifically to AI tool usage. They called it “AI brain fry” — a state characterized by mental fog, slower decision-making, increased errors, and a persistent difficulty focusing.

They emerge specifically from the supervisory demands of overseeing AI systems: monitoring outputs, evaluating quality, managing multiple parallel workflows, and making rapid context switches between them. Workers experiencing brain fry reported 39% more major errors and 39% higher intent to quit.

(For a more thorough analysis of the BCG study and related research, I recommend this post by Carl Hendrick.)

The New Multitasking

Coworking alongside AI agents and managing them is a strange dynamic that feels simultaneously synchronous and asynchronous.

Some tasks are performed instantaneously. But arguably, the real value from agentic work comes from the longer, more complex tasks. Those wait times are often unpredictable — you get a sense for which tasks will require more time, but it’s inexact. It may be a few minutes or it may be hours.

To realize the full productivity gains of agents, you don’t just want to stare at the thought traces while you wait — you’re moving on to the next task. And if you’re like us, that task involves opening up a new chat with Claude and getting it started on something else.

UC Berkeley researchers studied this pattern ethnographically over eight months at a tech company and found that AI didn't free up workers' time — it expanded what they felt capable of taking on. They called it "silent workload creep."

Cross-Chat Chaos

A real example from the chaotic way I work now, all happening in Claude at the same time across two computers:

1. Code: Writing a script to gather published text and assemble into a corpus of markdown files

2. Chat: Conducting a preliminary literature review

3. Cowork: Generating a rough outline of an article based on a Wispr Flow brain dump of ideas, using my custom skills and context files for article-writing

All three of these work streams are happening simultaneously, and are all related to a single project (writing an article). But while those are running, I also have some unrelated tasks:

4. Cowork: Gathering status updates from my direct reports, assembling them in an update message, and sending to the CEO

5. Chat: Diagnosing a weird malfunction happening with my “smart” thermostat

So now I have 5 different tasks, some related, some not, all requiring my attention, error-catching, approval, or redirection at various moments. I’m in a continuous state of intense multitasking. Each time I switch between parallel tasks, I have to reorient and remember which task I’m looking at, assess what the agent has done, and decide how to move it forward.

Just to add another meta layer on top of it all, I’m thinking about how to improve my reusable skills and workflows, and what else I can do to improve my agents’ efficiency. And that gives me an idea for another Substack article…😵‍💫

Psychological research has long documented that human beings are pretty bad at regular multitasking. Sophie Leroy's research on attention residue shows that switching away from an unfinished task degrades your performance on the next one — and every agent task you've kicked off and walked away from is, by definition, unfinished. Gloria Mark's research on interrupted work found that interrupted workers compensate by working faster, but at the cost of higher stress and more errors. Sound familiar?

This Is a Design Problem

While humans will likely always struggle with multitasking, the current design of agentic AI tools make this so much harder than it has to be.

Consider Claude Cowork — possibly the system the majority of consumers and knowledge workers are currently using as their personal / professional agentic AI. Compare its design to Claude Chat. Yes, there’s a slightly more informative sidebar with a linear to-do list and list of referenced files. But other than that, they’re nearly the same design.

Cowork is still designed for the kind of quick question-and-answer interactions of 2023. It has barely evolved for agentic AI.

I’d like to call out just a few particularly egregious design flaws that make my agentic multitasking harder.

1. Returns Me To the Wrong Place

Every time I switch between modes (Chat, Cowork, Code) or individual chats, Claude inexplicably returns me to the top of my conversation. Yes, there’s a little arrow at the bottom of the screen that will take me back to the bottom, where I can find what I actually need (the latest response from my agent.) But having to click that button every single time I switch between chats becomes tedious fast.

2. First-Prompt Thread Naming

Claude, like most consumer AI chats, generates a name for each thread based on the very first prompt it’s given. That works fine for a simple information-seeking or email-generating task, where each chat is focused on a single objective. It does not work for the kind of long, complex, meandering conversations you have during agentic coworking tasks.

This is especially problematic when those conversations (as often do) naturally evolve into multiple branching tasks related to the original request. What started as a literature review for a specific article idea may naturally develop into new research questions and multiple articles.

3. No Support for Branching

Claude just isn’t designed for that type of branching. the clearest example is the linear to-do list that Cowork generates to keep the user updated about its plan and progress. In my experience, that to-do list isn’t very useful, because it doesn’t reflect my thinking or Claude’s actual workflow. In reality, Claude is often spinning up multiple agents to branch out, work in parallel, and recombine outputs. A linear chat simply does not work for nonlinear thinking.

When switching between multiple related and unrelated tasks, Claude (and most other major AI chat products) makes things harder than they need to be. To name a few: starting at the top for each thread, a linear (and wordy) progress tracker, and labeling chat threads from the first prompt.

This is why canvas-based agentic tools like Eden are attractive — they’re better aligned with the way human beings actually think, synthesize, and produce.

Eden’s approach to agentic coworking is nonlinear and canvas based.

There’s more to unpack here. Sarah and I are working through a more detailed analysis. But I can tell you right now, one of the core problems is that most AI companies are focused on making small, data-informed tweaks to their interface, rather than grounding their design strategy in a deeper understanding of user psychology and behaviors.

Are you feeling this too? What (if anything) helps you work around these design limitations to feel a little less scattered?

References

Bedard, J., et al. (2026). When Using AI Leads to “Brain Fry.” Harvard Business Review / BCG.

Hendrick, C. (2026). Referenced Substack post.

Leroy, S. (2009). 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, 109(2), 168–181.

Mark, G., Gudith, D., & Klocke, U. (2008). The Cost of Interrupted Work: More Speed and More Stress. CHI 2008.

Osmani, A. (2026). The 80% Problem in Agentic Coding. Substack.

Ranganathan, A. & Ye, X. M. (2026). AI Doesn’t Reduce Work — It Intensifies It. Harvard Business Review.