It sounds like AI agents should leave us rested. They take routine work off our hands, so we should have more time and more energy. In practice, I hear the opposite more and more often: “I used to sit through the whole workday. Now by lunch I can only stare at the wall.”
Personally, I think the explanation comes from the agents themselves. Let’s look at a person the way we look at an agent: as something with a context window.
One domain, one window
Look at how work used to be. A developer adds a button to a website. Their head is busy with one domain: the interface, the programming language, the framework. A good manager knew this and gave each person tasks where they were strong and fast.
Now an agent adds the button. The person only gives the task. Meanwhile, other agents write articles, draw designs, make videos, and run research. Instead of a frontend developer, a backend developer and a product manager, we have one product builder who orchestrates agents: marketing, product, development, hotfixes, calls with the business owners. And then someone sends a personal question in the chat.
Researchers at UC Berkeley saw the same at a tech company that gave everyone AI tools: people took on tasks that used to belong to others (Berkeley Haas).
From a city to continents
Imagine a map. Before, a person’s brain moved around one city. A senior specialist moved around a country. Now the brain travels between continents several times a day, and every trip costs far more than a ride across town. Is it just more work? No. It is much more expensive work.
There is a second problem. In our home domain, the roads are built: we have walked them a thousand times. Between domains, the brain often walks a new path. That takes time and energy, and by the evening the head hurts.
Research points the same way. In experiments by Rubinstein, Meyer and Evans, switching between tasks cost time, and the cost grew when the tasks were more complex or less familiar (APA). You may have seen the figure “up to 40% of productive time”: that is Meyer’s own estimate, not a result of the experiments. Sophie Leroy described attention residue: after a switch, part of our attention stays with the previous task (Leroy). So when you fly between continents, part of your luggage stays behind.
Yes, some people have always lived like this. Entrepreneurs, for example. But it is a small group, and it is hard for them too.
Everyone is running ten times faster
The second reason is social. If agents do the work, why wait a week? Stakeholders now expect a result after lunch, in an hour, at worst by the evening. Before, you had a week to sit and think. Now there is a rush, pressure, and a constant “faster, faster”.
The expectations are measurable. In a 2024 Upwork Research Institute survey, 96% of executives expected AI to raise productivity, while 77% of employees who used AI said it had added to their workload (Workplace Intelligence). Upwork sells freelance work, so take it with a grain of salt. The direction still matches what I see.
We do it to ourselves too. Agents are fast, so we want everything done now: skip the break, skip lunch, stay until night. We do less of the work with our own hands and sit at the computer longer. Anthropic’s engineers estimated that 27% of their work with Claude would not have been done at all without it (Anthropic). That is not saved time. That is new work.
The myth of the infinite context window
Put this together, and you get a new belief: if an agent can jump between domains, a person can too, so you can load into a person’s head as much as you like. Is that true? No.
Nobody used to ask one developer, on one day, to add a button, write a marketing plan, go to the bank and discuss the business with stakeholders. Everyone knew it was too much. Now it is called a normal day.
You don’t have auto-compact
Here the agent gives us the best analogy. When an agent’s context fills up, it starts to lose track of earlier instructions. Claude Code then compacts the conversation, and if you don’t say what to keep, something valuable can disappear. That is why you can write /compact focus on the auth bug fix. And an automatic compaction in the middle of a task comes at the worst possible moment (when to compact).
People don’t have a /compact command. Our compaction happens when we rest: when we sleep, walk, meditate, do something with our hands. Sleep is the big one. In deep sleep, slow waves of cerebrospinal fluid wash through the brain; researchers saw it in sleeping people with fast MRI in 2019 (Fultz and colleagues, Science). One sleepless night raised beta-amyloid, a protein linked to Alzheimer’s disease, by about 5% in the brains of 20 healthy people (NIH). And while we sleep, the brain is busy sorting what we learned during the day and fixing it in memory, as Diekelmann and Born summarized in their 2010 review. Even ten minutes of quiet rest with closed eyes after learning something helped people remember it a week later, in a small study of older adults (Dewar and colleagues, 2012).
If nobody gives you that time, the window overflows, and you stop taking in information, just like an agent with a full context. The human version of an auto-compact in the middle of a task is called staring at the wall after lunch.
A chainsaw doesn’t mean 101 baobabs
Before, a man cut down a big baobab with an axe in half a day. Now he has a chainsaw and does it in 30 minutes. That doesn’t mean he should spend eight hours running around and cutting down 101 baobabs.
Some things will get faster. The economy will adjust, and the standards will change. But a person is still a person. That is the important part. People who studied, created or worked hard all day got headaches long before AI, and they used techniques that protect mental energy.
What still works
New techniques for the agent era haven’t been invented yet, and I’m not sure they need to be. The old ones still work, because the person hasn’t changed.
-
Focus blocks and real breaks. Pomodoro is decades old and still the simplest way to alternate effort and recovery. With agents, 25/5 and 50/10 both work; what matters is that the break is a break (Pomodoro with AI coding agents).
-
One domain per block. Code in one block, marketing in the next. Fewer continents per hour means fewer expensive trips.
-
Rest that holds your attention gently. A wandering mind is not a resting one: in a study of 2,250 people, minds wandered almost half of waking time, and people were less happy while it happened (Harvard Gazette). A walk in a park restored attention better than a walk down busy streets (APS). In a small study, experienced meditators’ minds wandered less (PubMed). Scrolling a feed is not on this list.
-
Sleep. It is the only full compaction we have.
-
Honest deadlines. Agents are fast; review, decisions, and people are not. Say it out loud when you agree on a deadline.
What I think
There is a small chance that you push through and achieve something big this way. There is a bigger chance that you burn out and give up. Neither is a plan. We have to learn to live and work with agents for years, not for one sprint.
So try an experiment. Take a real break. Come back and look at what happened while you were away. I promise nobody died in the meantime, and at the end of the day you will feel less tired.
The tools got faster; you don’t have to. Take one real break tomorrow and tell me what happened.
