Over the last year, I noticed something about the way I work with AI.
I get much more done, but I also feel much more tired. Not “I need some coffee” tired. I mean completely mentally drained at the end of the workday, especially by the end of the week.
The strange part is that I am not working more hours. AI lets me do in one hour what could easily take half a day before.
That hour is just much more intense than it used to be.
A normal developer workday was never eight hours of nonstop hard thinking. You wrote code, read docs, waited for a build, got stuck on a problem, went for coffee, came back, talked to someone, tried something else.
The mental load was naturally spread across the day.
With AI, this structure starts to disappear.
You can have Claude Code open while another agent is refactoring something, a third one is working on tests, and a fourth one is researching a problem. While one is working, you start another.
Then the first one comes back with a result. You need to read it, understand it, check it, make a decision, and give the next prompt. While you are doing that, the second one finishes. Then the third.
You are no longer just writing code. You are running a small factory of mental work.
And the factory scales much better than the person running it.
Humans have usage limits too
I like the comparison with Claude Code limits.
Claude has a five-hour usage window. This does not mean “five hours of tokens” in a literal sense. Usage depends on the length and difficulty of requests, the model, tools, and other things.
If I start more work in parallel, I do not get endless compute. I can simply burn through the available budget faster.
Parallel subagents = faster AI budget consumption.
I think humans work in a similar way.
We do not have a real token counter in our heads, but we do have limits. Attention and working memory are limited. So is our ability to make difficult decisions well, switch between several complex problems, and keep a large system in our head at the same time.
Parallel tasks = faster cognitive energy consumption.
AI does not increase this resource. It increases how fast we can spend it.
Before AI, I could spread a lot of difficult thinking across an eight-hour workday. Now I can burn a large part of that mental budget in two or three very intense hours with several agents.
I got more compute around my brain.
But I did not get another brain.
AI removed the waiting, not the thinking
This is the part I think we miss when we talk about AI productivity.
AI removes many slow parts of software development: writing boilerplate, searching docs, trying basic implementations, refactoring repetitive code, or manually working through simple tasks.
That sounds like it should make the workday easier.
But it also removes a lot of natural low-intensity time.
Before, there were gaps. Now the next piece of work can arrive almost immediately.
Finished?
Here is another result.
Checked it?
Here is the next one.
Made a decision?
Start three more agents.
The machine does not get mentally tired. So if you let it, it can keep feeding work into the human bottleneck all day.
There is already research pointing in this direction.
A 2026 Harvard Business Review article is literally called “AI Doesn’t Reduce Work—It Intensifies It”.
Researchers from UC Berkeley followed employees at a tech company for eight months. Instead of simply finishing the same work faster and getting more free time, people started working faster, taking on more tasks, and spreading work across more of the day.
Often, this happened without direct pressure from managers.
AI created more room for work, and work filled it.
Software development is shifting from creation to verification
A 2026 study about software engineers found something similar: The Impact of AI Coding Assistants on Software Engineering: A Longitudinal Study.
The authors found a shift from creation to verification.
82% of participants said they spent less time directly writing code. But another type of work grew. The authors called it supervisory engineering work: giving tasks to AI, checking its output, fixing mistakes, and deciding what should happen next.
84% of developers said their productivity improved. At the same time, among matched participants, the share of people who reported a worse developer experience in at least one area grew from 14% to 27%.
The problems included flow state and cognitive load.
This describes what I see pretty well.
AI removes some of the work needed to create the result, but it creates much more work around checking the result.
And checking also uses your brain.
Sometimes it is harder than writing the code yourself, because you need to understand enough to know when the AI is wrong.
AI brain fry
In March 2026, HBR published another article: When Using AI Leads to “Brain Fry”.
It was based on research involving 1,488 workers. The study connects heavy AI oversight and higher workload with problems focusing, slower decisions, mental fatigue, headaches, decision fatigue, and a higher chance of mistakes.
So we have reached a strange point.
AI can produce work faster than a human can properly process it.
The eight-hour workday starts to look strange
Most office work is still measured in time.
You have an eight-hour contract. But eight hours of being at work and eight hours of hard reasoning are completely different things.
A person can physically sit at a computer for eight hours. That does not mean they can spend eight hours making good architecture decisions, keeping several complex contexts in their head, and checking the results of five agents.
Before AI, a workday naturally had different levels of intensity. Some parts were hard. Some were routine. Some were waiting.
AI can compress much more difficult work into the same block of time.
And this is where the eight-hour model starts to look strange.
If I can spend most of my mental energy by early afternoon, the clock may say I still have five hours left.
My brain does not care.
Who gets the extra productivity?
This is the part I find uncomfortable.
My contract still buys roughly the same amount of my time. My salary does not automatically grow because Claude Code can now help me produce much more inside that time.
But the amount of work that can fit into those hours grows.
So who gets the extra productivity?
We built technology that could reduce part of the mental work people have to do.
Then we did the most predictable thing possible.
We filled the new free capacity with more work.
Not:
Finish the same work faster and stop earlier.
But:
You can do more now, so do more.
After a while, the higher output stops looking exceptional.
It becomes the baseline.
The takeaway
I am not against AI coding tools. Quite the opposite. I do not want to go back to software development without them.
But productivity and human capacity are two different things.
If AI lets me spend in two hours the mental energy that used to be spread across eight hours, the answer cannot be to keep the same level of intensity for another six.
We understand this perfectly when the limit belongs to a machine.
Claude has usage limits. APIs have token limits. CPUs have thermal limits.
Run more work in parallel and you hit those limits faster.
But when the limit belongs to a human, we still look at the clock and say:
You have six hours left.
That makes less and less sense to me.
AI scales execution. It scales parallel work. It scales output.
It does not scale the human brain.
Otherwise, the great AI productivity revolution may end up doing something much simpler:
teaching us how to burn through a full day of human cognitive energy before lunch.
Top comments (0)