Are AI agents frying power users brains? The flip side of turbo productivity
Anyone living on code, sprints, and continuous deployment has already noticed: the arrival of AI agents changed the game. Tools capable of writing, testing, and shipping code almost autonomously have become the new standard for high-performance teams. But behind the promise of infinite productivity, a heavy side effect is starting to show up: mental fatigue, prompt addiction, and a very real feeling of a fried brain.
The landscape that is emerging is both curious and concerning. Senior developers, founders, and tech leaders report marathon agent sessions, 16, 18, 19-hour work cycles, fear of wasting paid tokens, and real difficulty turning their minds off at night. Instead of just speeding up the coding flow, agents are shaking up how these people think, work, and rest.
AI agents in practice: from copilot to agent swarms
Early code assistants seemed harmless: they suggested line by line, helped with a refactor, handled some boilerplate. Now, the story is different. Agentic systems from players like Anthropic, OpenAI, and open-source projects can:
- understand the context of an entire repo
- write full code modules
- run automated tests
- fix errors and try again
- open PRs and even describe what changed
In practice, the devs role shifts from direct author of the code to something closer to a swarm-of-agents orchestrator. The workflow becomes a continuous loop:
- craft the prompt
- let the agent act
- review the result
- correct the course with a new prompt
- repeat that all day long
This dynamic looks amazing if you only measure output. Less manual code, more features, less time. But under the hood, there is a growing cognitive cost: the brain is stuck in constant supervisor mode, evaluating responses, adjusting context, and trying to track multiple parallel processes at once.
When productivity turns into AI psychosis
Some of the most well-known names in AI and startups are starting to put into words what they are feeling in this new environment. And the stories are not light.
Andrej Karpathy and the AI psychosis state
Andrej Karpathy, cofounder of a major AI big tech company and one of the people who helped popularize the concept of vibe coding, said on a podcast that he has been in a state he calls AI psychosis since December. What does that mean in practice?
- before, he wrote about 80% of the code by hand and delegated 20% to AI
- very quickly, that flipped to 0/100: practically all the code was handed off to agents
- he describes routines of up to 16 hours a day commanding agent swarms, issuing instructions, evaluating responses, and diving deeper into possibilities
One detail stands out: he pays a monthly subscription to use these systems and, when he notices the month is ending with tokens left over, he reports feeling strong anxiety. The sense is that if he does not use everything, he will fall behind compared to people squeezing every cent of compute and every prompt.
This mix of pressure to maximize tool usage with the feeling of exploring a new, almost limitless territory creates the perfect environment for excess: more hours online, less sleep, more experimentation loops.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Garry Tan and cyber psychosis
Garry Tan, CEO of Y Combinator, also jumped into the conversation with his own term: cyber psychosis. He has shared that he stayed awake for 19 hours straight playing with code tools and only went to sleep after 5 a.m. This marathon is not a one-off: it is part of an intense phase of exploring the limits of AI in software development.
The curious part is that when he saw other founders and CTOs bragging about 36 hours without sleep to ride alongside their agents, he made a point of calling it out: that is clearly not healthy, including based on his own experience. In other words, the people on the front lines feel both the upside and the cost of this work style.
Cognitive limits and the infamous brain fry
Beyond product and engineering folks, researchers and veteran devs are also starting to flag something more structural: it is not just tiredness, it is a specific kind of cognitive fatigue tied to heavy AI use.
Simon Willison: blowing the mental stack
Simon Willison, a developer with more than 25 years of coding experience before the era of generative AI, describes the phenomenon in a simple way: there is a clear limit to how much a person can hold in their head at once. With agents, that limit is easily blown past.
He talks about something like blowing the minds stack: too much information, too many contexts, too many parallel flows. Constantly supervising what agents are doing demands a level of attention that the brain cannot sustain for long without paying a price. And he is blunt: trading sleep for marathons of agentic coding is, in practice, an obvious road to exhaustion.
Tim Dettmers: max productivity, minimal focus
Tim Dettmers, an AI researcher at a major research institution, points to another key aspect: the biggest productivity gains show up when you use multiple agents in parallel. But that pushes humans into a kind of work very few people can handle well: constant context switching.
Instead of going deep into a single problem at a time, the dev ends up:
- monitoring several agent flows in parallel
- switching tasks every few minutes
- keeping the state of multiple reasoning threads in their head
This need to hop between mental tracks does not match how the human brain works best. Dettmers himself sums up the paradox: agents expand what seems possible, but they also amplify the stress on focus and mental capacity.
From excitement to pathology: when it becomes addiction
Little by little, behavior around agents is starting to look less like a nerdy hobby and more like a addiction pattern. And that is not just a metaphor.
Gamification, reinforcement, and the slot machine effect
Simon Willison and other devs highlight dynamics that look a lot like betting mechanics in how these tools are used. Armin Ronacher, another well-known name in the community, summed it up neatly: lots of people have been hit by a kind of agent coding addiction. It feels good, the output is amazing, sleep disappears.
Quentin Rousseau, CTO and cofounder of an incident platform, broke this effect down in detail. After moving his workflow to a strongly agentic model, he spent months unable to sleep properly. His brain just would not shut down. He even had to turn to doctor-prescribed medication to get any sleep.
He says that while he is a supporter of AI acceleration, agents end up working like slot machines:
- you fire off a prompt
- a quick response comes back
- part of the code is done
- that releases dopamine and reinforces the behavior
- you immediately go back for one more prompt
Sometimes the agent fails badly, delivering something broken or totally off. And it is exactly that mix of amazing hits with chaotic misses that mirrors variable reward systems, very common in gambling. It keeps the brain stuck in the just one more try loop.
Founders as the first collateral damage
Rousseau also points out that founders and tech leaders tend to be even more vulnerable to this pattern. They are already naturally performance-driven, want to squeeze the most out of their tools, and usually operate under massive time and delivery pressure.
When you combine:
- agents that promise crazy acceleration
- a culture of extreme productivity
- fear of falling behind
you get the perfect breeding ground for addiction to productivity tools. In his words, these people end up being the first collateral damage of systems that are, in part, designed to be highly engaging.
Brain fry: the cognitive cost of over-supervising AI
Researchers at major consultancies and universities are already giving this effect a more technical label: brain fry. In practice, it is a form of mental fatigue that shows up when someone pushes past their cognitive capacity limit by using or supervising AI systems for too long.
A study published in a top business journal looked at this phenomenon in corporate settings and surfaced some heavy findings:
- the mental effort required to work with AI increases human errors
- decision fatigue also goes up, that feeling of being too drained to choose what to do next
- there is a higher intention to quit after a while at this pace
One interesting point is that many companies still use token consumption, hours of use, and volume of AI interaction as a productivity proxy. In practice, that encourages heavy usage patterns, without accounting for the cumulative cognitive load on people supervising these systems.
Has this always existed or did AI raise the stakes?
To be fair: devs locked in a room, pulling all-nighters and crashing on the office couch are nothing new. Stories of crunch marathons, insane deadlines, and teams basically living at work go way back before agents. Even at big tech companies, there have been cases of teams sleeping on factory floors or in their chairs to hit targets.
The difference now is the pace and continuity. Before, even during crunch time, there was a hard physical limit: your hands get tired of typing, the mind burns out more slowly when the flow depends on manual output. With agents running 24/7, ready to respond in seconds, the brain is stimulated nonstop. It is easy to fall into the trap of just one more prompt, just one more test, just one more round.
Work, which was already intense, picks up an extra layer of pressure: it is no longer just the deadline racing; now it is also the machine that never stops and that, in theory, could be producing more while you are trying to rest.
Balancing agents, focus, and mental health
Amid all this, one thing is clear: AI agents are not going anywhere. On the contrary, they are likely to become more powerful, more autonomous, and more deeply integrated into development, operations, and product stacks.
The challenge is not choosing whether to use them or not, but figuring out how to:
- recognize individual cognitive limits
- design processes that avoid endless supervision without breaks
- separate intense exploration phases from real rest periods
- reduce the tight link between extreme token usage and professional success
Experienced devs like Simon Willison are already spelling it out: each person needs to understand their own limits and build routines that steer away from the obvious burnout path. Coding with agents instead of sleeping, over and over, might work for a few days but will not hold up.
The big irony is that tools built to boost our capabilities can, if used without guardrails, blow right past the very thing that lets us think clearly: our mental bandwidth. And if current power-user behavior is any sign of what is coming, the brain fry that is now hitting devs and founders could easily spread to any field that embraces AI agents as the main work engine.
The next chapters of this story will probably focus less on what agents can technically do and more on a very human question: how far is it worth going to keep up with the machines speed without letting your brain melt in the process?
