Token Anxiety: Why Developers Are Burning Out Faster in the AI Era

Leaving the Party at 9:30 to Check on Your Agents

In February 2026, software engineer Nikunj Kothari published an essay called "Token Anxiety" that hit a nerve across the tech industry. It opens with a scene that has become disturbingly common in San Francisco:

"A friend left a party at 9:30 on a Saturday. Not tired. Not sick. He wanted to get back to his agents. Nobody questions it anymore."

The essay describes a cultural shift where developers check what their agents produced overnight before coffee or texts. Saturdays have become uninterrupted build windows. Parties are sober now because attendees want to stay sharp for the code they will write when they get home. Social conversations have moved from "what are you building?" to "how many agents do you have running?"

Y Combinator CEO Garry Tan captured the vibe in a single tweet: "I'm giving up drinking because of Claude Code. I need my brain to be maximally pristine so I can sling 10k LOC a day."

This is not hustle culture rebranded. This is something different. And it is spreading fast.

The Paradox Nobody Expected

AI coding tools were supposed to reduce developer burnout. The pitch was simple: let AI handle the boring parts, and developers can focus on the interesting problems. Less grunt work, more creativity, better work-life balance.

The opposite is happening.

A Bloomberg investigation in February 2026 reported that AI coding agents like Claude Code are "fueling a productivity panic in tech." UC Berkeley researchers published findings showing that AI is accelerating work, not reducing it. Productivity gains are morphing into burnout and workload creep.

The numbers tell the story. According to research compiled in early 2026, 95-98% of developers believe AI tools could reduce burnout. But in practice, 67% report spending more time debugging AI-generated code, and 68% spend more time fixing AI-related security issues. The tool that was supposed to save time is creating new categories of work.

As one developer wrote in a Medium post that went viral: "My efficiency increased 10x, yet I'm more exhausted than ever."

The Faustian Bargain of Vibe Coding

ThePrimeagen, one of the most respected voices in the developer community, recently shared his own experience with this phenomenon. After years of skepticism toward AI coding tools, he decided to force himself into the vibe coding workflow to understand what people were experiencing.

What he discovered was a trap disguised as productivity.

The initial phase feels amazing. You describe what you want. The AI builds an MVP in minutes. Ideas that would have taken days to prototype are up and running before lunch. ThePrimeagen built stream control tools, meme managers, and OBS automation, all through vibe coding.

But then the second phase hits. Each fix requires more prompting, more waiting. The cycle is long. And because you can spin up new projects so quickly, you end up with three or four half-built things running simultaneously. All of them in the "most crappy versions of themselves," with the work growing at an exponential rate.

His insight cuts to the core of the problem: "The programming part was never the problem. The problem was solving the right problem."

Before AI, the difficulty of implementation acted as a natural filter. You could not try every idea, so you had to be selective. You picked one project and committed. Now you can have four agents spinning at once, all producing mediocre code, all demanding your attention, all creating more work than they finish.

10,000 lines of code in a day. Easy. But that does not mean those 10,000 lines were any good. It does not mean you feel satisfied. It does not even mean you built the right thing.

The Anxiety Is Rational, Which Is Why It Sticks

The Token Anxiety essay nails why this feeling is so hard to shake: the anxiety is rational.

Every week, some new benchmark drops that makes last month's workflow feel prehistoric. New models ship overnight. Context windows double. None of this reduces the pressure. It multiplies it. You can do more now, and someone already is.

The window to be first at anything feels like it is shrinking by the day.

This creates a specific type of stress that traditional burnout advice does not address. "Take a break" does not work when your competitor's agents are running while you sleep. "Set boundaries" does not work when the CEO of the most prestigious startup accelerator in the world is publicly bragging about slinging 10,000 lines of code per day.

The result is what researchers are calling a new form of occupational anxiety. Not the old "I have too much work" burnout. Something more insidious: the feeling that you are falling behind every moment you are not producing.

What the Research Actually Shows

Harvard Business Review published findings showing that AI is accelerating work, not reducing it. The productivity gains are real, but organizations treat every minute saved as a minute available for more work.

A Fortune report on UC Berkeley research found that employees using AI tools increased both the work they could complete and the variety of tasks they could tackle. But as one worker put it, "I don't work less. I work the same amount or even more."

The context-switching problem is particularly brutal. A developer might touch six different problems in a day because "each one only takes an hour with AI." But context-switching between six problems is expensive for the human brain. The AI does not get tired. The developer does.

And here is the part that should concern every engineering manager: the developers most likely to burn out are the ones who embraced AI the hardest. The high performers. The early adopters. The ones who are supposed to be thriving.

The $1,000-a-Day Token Habit

According to developers in the community, there are people in the Bay Area spending $1,000 a day on API tokens. They have multiple agents running around the clock. They have dashboards tracking agent outputs. They have ideas flowing through their heads faster than any human could execute.

And many of them have literally nothing to show for it.

The work is multiplying. The task juggling is getting out of control. The productivity metrics are through the roof. Yet nothing meaningful ships. Because the bottleneck was never writing code. The bottleneck was always deciding what to build, why to build it, and whether it actually solves a real problem.

AI removed the implementation friction. But implementation friction was never the hard part. The hard part was product thinking, architectural judgment, and the discipline to finish one thing before starting three more.

What Nobody Wants to Hear

The uncomfortable truth is that most of what these agents produce does not matter. Not because the code is bad (though often it is). But because the projects themselves do not solve problems anyone is willing to pay for.

The developers who are actually winning in 2026 are not the ones with the most agents running. They are the ones who:

1. Pick one problem and solve it completely instead of spinning up four half-baked prototypes
2. Use AI selectively for the tasks it is genuinely good at (boilerplate, first drafts, explanations) instead of trying to vibe-code entire products
3. Maintain their own judgment about architecture, design, and product decisions instead of outsourcing their thinking to a model
4. Set actual boundaries even when the culture tells them they should be grinding 24/7
5. Ship finished products instead of accumulating an ever-growing list of 80%-done projects

The Historical Pattern

This is not the first time a productivity revolution created more stress instead of less.

When email replaced letters, the promise was faster communication with less overhead. Instead, we got an average of 121 emails per day and the expectation of instant responses. When smartphones gave us mobile access, the promise was flexibility. Instead, we got the expectation that we are always available.

Every tool that makes work faster eventually gets absorbed into the baseline expectation. The new floor becomes the old ceiling. And the humans using those tools end up working more, not less.

AI coding tools are following the exact same pattern. Today's "10x productivity" becomes tomorrow's minimum expectation. The developers who burned through their weekends to ship faster will find that their weekend pace is now the weekday standard.

What Actually Helps

If you recognize yourself in any of this, here is what the research and the developers who have been through it suggest:

Separate building from deciding. Spend time without any AI tools deciding what is worth building. Then use AI to build it. Do not let the ease of starting something substitute for the judgment of whether it should exist.

Limit your active projects. Having four agents running does not make you four times as productive. It makes you a project manager for four mediocre codebases. Pick one. Finish it.

Accept that someone will always be "ahead." The person spending $1,000/day on tokens is not your competition. Your competition is the version of you that ships something complete and useful.

Protect your non-coding time. ThePrimeagen's advice is worth repeating: "That one extra feature in your calendar app is not worth skipping out on some good times with your friends. Hard work got me to where I am now, but it is not who I am."

Remember what the tools are for. AI is a powerful assistant. It is not a replacement for taste, judgment, or the ability to decide what matters. The developers who treat it as a thinking replacement will burn out. The ones who treat it as a building accelerator will thrive.

The most productive thing you can do in 2026 might be closing your laptop at 9:30 and staying at the party.