Andrej Karpathy, one of the most respected names in AI and a former Tesla AI director, says he hasn't personally written a single line of code since December 2025. On a recent episode of the No Priors podcast, he described going from writing roughly 80% of his own code to zero - replaced entirely by directing AI coding agents.
He calls the state he's in "perpetual AI psychosis." Not a clinical term, but a vivid one: the feeling that the possibilities are so vast, and the agents so capable, that you can't stop. Sixteen-hour days spent not coding, but orchestrating AI that codes for you.
He's not alone. Y Combinator president Garry Tan used the phrase "cyber psychosis" to describe the same phenomenon - sleeping four hours a night because he can't stop building with Claude Code.
The Shift From Writing to Directing
This is a meaningful data point because of who's saying it. Karpathy isn't a weekend hobbyist experimenting with Copilot autocomplete. He's someone who could write neural network training loops from scratch and previously did so daily. When someone at that skill level stops writing code entirely, it says something about where AI coding tools have landed.
The pattern he describes - going from "AI assists me" to "I assist AI" - is showing up across developer communities. People running multiple Claude Code sessions in parallel, treating each one like a junior developer that needs a clear task description and periodic check-ins. The skill that matters isn't syntax or algorithms anymore. It's decomposing problems into agent-sized chunks and reviewing output quickly enough to keep the pipeline moving.
The Burnout Question Nobody's Asking
There's a less glamorous side to this that deserves attention. Working 16-hour days directing agents isn't a productivity breakthrough - it's a recipe for burnout with extra steps. The "psychosis" framing, even used half-jokingly, points at something real: when your tools can work faster than you can think, the bottleneck becomes your own judgment and attention span. And those degrade badly on four hours of sleep.
The productivity gains from AI coding agents are genuine. A task that took a day can sometimes finish in an hour. But the response many developers are having isn't to reclaim those hours - it's to fill them with more agent sessions. The work expands to consume the capacity.
For everyday users of tools like Claude Code, Cursor, or Aider, the practical takeaway is more grounded than the "psychosis" framing suggests. These tools are genuinely changing what a single developer can build. A solo founder can now ship features that previously required a small team. But the developers getting the most sustainable value aren't the ones sleeping four hours - they're the ones who use agent sessions to finish their actual work faster, then close the laptop.
Karpathy's experience is real and worth paying attention to. The question isn't whether AI coding agents are powerful enough to replace manual coding for experienced developers - that's clearly settled. The question is whether the humans directing them can maintain the judgment that makes the output worth shipping.