Something rots in the developer brain, and nobody wants to name it. I am naming it: Token Anxiety. You wrote code in the last six months, so you felt it, a low hum behind your eyeballs, the itch to check a running agent while someone talks to you at dinner. A job could run right now, and it stays idle while you feel the blame anyway.
A Buzz You Used to Earn
Every programmer who codes long enough knows this feeling. You hit a wall, a bug or an architectural knot that defeats your conscious mind, so you step away, take a shower, or go for a walk.
Then the solution lands somewhere between the shampoo and the conditioner. The whole architecture clicks into place in one white flash. That was the old buzz. It felt like a superpower.
The dark side was subtle. You sat at dinner, present in body and spiraling in your head, and the food went cold while half of you sat inside a function signature. It was intimate, and in a sick way it was almost romantic. But it came and went, because you had to earn it, and a hard problem had to show up first.
The buzz is permanent now, and it comes from the machines instead of your brain.
Leaving the Party at 9:30
A culture takes shape among developers right now, and it unsettles me.
Someone leaves a party at 9:30 on a Saturday. Tiredness is not the reason, and illness is not the reason. They want to get back to their agents. Nobody questions the exit, because half the room wants the same thing, and the other half checks agent progress on a phone at the party.
Andrej Karpathy, the OpenAI co-founder who coined the term "vibe coding," told the "No Priors" podcast that he had been in a "state of AI psychosis" since December. Axios reported the story in April 2026. He spends 16 hours a day issuing commands to agent swarms now. His ratio of hand-written to AI-delegated code flipped from 80/20 to 0/100 that same December.
He pays for a monthly token subscription, and when he has tokens left over near the end of the month, he says he feels "extremely nervous." He rushes to burn through them so he does not fall behind everyone else. He is describing anxiety about unused compute in the same terms other people use for anxiety about an unread inbox.
Karpathy describes the AI psychosis and the leftover-token nerves in this No Priors episode.
A novel feels indulgent. A movie with no laptop open feels wasteful. One voice in your head says that something could run right now, and the voice never shuts off. That is Token Anxiety, and the frightening part is that it is rational, which is exactly why it sticks.
Every Faster Model Raises the Stakes
Every week a new benchmark makes last month's workflow look prehistoric. A new model ships overnight. Context windows double. Claude Opus gets faster. Claude Sonnet gets cheaper.
None of this reduces the pressure. It multiplies the pressure. You can do more now, and somebody already does it, so the window to build first, or to ship first, shrinks by the day.
Tech circles once felt like a raw electric field of opportunity. The message ran "there is so much to build, come try things," and the optimism was genuine, infectious, and reckless.
Look at the same circles now. The energy inverted. The message no longer says "come build exciting things." It says "I better keep working, because one hour of rest costs me everything." The tone moved from opportunity to desperation, from "I get to build" to "I must build or I drown," and that is a different problem, one that eats people alive.
Vibe Coding's Faustian Bargain
I forced myself into the vibe coding world. I build things with agents, and I start projects through conversational prompts, because I wanted to understand the hype from the inside instead of arguing about it from a distance.
The beginning of every project seduces you. The idea is simple, you describe it in a few words, and the rough scaffold comes up fast. You get a minimum viable product, or MVP, in hours. The speed feels electric.
Then the fixes start. Each fix needs more prompting, more waiting, and more context-loading. The cycle runs longer than a Rust compile, and each fix spawns two more tweaks you want, and those spawn two more.
Here is the insidious part. Each new project starts so easily that you soon run four or five of them at once, and every one of them sits in its crappiest version while demanding your attention. You babysit small steps and type explanations in English to fix work that takes ten keystrokes in real code.
You build more than you ever built in your life, and you accomplish less than ever. Paul Graham posted on X about a founder who told him he now writes 10,000 lines of code a day thanks to AI. Graham called it "probably the limit case." Graham was careful to note the founder is a hotshot who knows the tools well and that the code was not "bug-filled crap." Most people who chase that number are not that founder. They write 10,000 lines with no effort and no satisfaction, having never checked whether it was the right 10,000 lines to write.
Choosing What to Build Was Always the Hard Part
Nobody talks about this part.
Before AI, you could not try every idea. You had to select. You picked one problem, committed to it, and pushed through it. That constraint was a gift, because it forced you to prioritize, and it forced you to solve the right problem, since you could not afford to solve the wrong one.
Now you run three or four ideas at the same time. Every one of them stays half-baked and grows in scope at an exponential rate. The work multiplies, the productivity metrics climb, but nothing ships.
METR measured the gap between how fast developers felt and how fast they were in early 2025.
Futurism reported that Marty Kausas, the 28-year-old founder of Pylon, works up to 92 hours a week and eats prepackaged meals from Bryan Johnson's Blueprint so he never has to think about food. Emily Yuan co-founded the AI finance company Corgi at 23. She put it directly to the same publication: "Why would I go drink at a bar if I can be building a company?" Both examples are less about writing bad code than about a generation that mistakes constant motion for a finished product.
The programming was never the bottleneck. The bottleneck was always the choice of what to build. AI removed the wrong constraint. It made execution trivially cheap and did nothing about the decision that actually matters.
One Foot on the Platform, One Foot on the Train
A fracture opens inside every developer who straddles the old world and the new one. You hand-code something, and halfway through a function one thought creeps in: I could vibe code this in thirty seconds. So you start an agent, and now you stand in two worlds at once.
You feel the pull of the agents all day. Better start a quick one. Let me prompt this side feature. That buzzing sensation grows and compounds into anxious energy wearing the mask of productivity, the energy of someone who knows that work could happen and does not happen.
What This Actually Costs
I lived in that state before AI arrived. I pursued building without a stop, stayed up through entire nights to program, and felt anxious at social gatherings because part of my brain wanted the editor back. Every minute away from code felt like a wasted minute, and it damaged friendships. It turned me into a process instead of a person.
AI agents bring that exact feeling back, louder now, because the machine never sleeps. There is always another prompt to send and another project to start. The guilt of rest is infinite, because the capacity for work is infinite.
That is an addiction in a hoodie, not productivity.
What I'd Tell Myself Now
One extra feature in your calendar app does not pay for a good evening with your friends. One more overnight agent run does not pay for the fog you carry into tomorrow's conversation with someone who matters. The 10,001st line of code today rarely changes your life. The moment you miss tonight might.
Hard work is a vehicle. It carried you here, but it is not who you are. Confuse the vehicle with the identity, and you stop being a person and start being a process. A process does not have dinner with friends. It does not read a novel or watch a movie and feel something.
You are not a process. The machines run without you, so let them run. Turn off the terminal, close the laptop, and leave the agents unsupervised for one evening.
The code stays there in the morning. The same bugs stay, the same half-baked MVPs stay, and the same growing task lists stay.
The people around you will not always stay.
The pressure behind token anxiety is genuine. The thing you are racing against is only a feeling, and a feeling does not ship a product.
Ship less. Ship the work that matters.
Related reading: AI Costs Part 1, Claude Code Memory Architecture
