Token anxiety

It starts small. You close your laptop for the night, and somewhere in the back of your head a voice whispers: something could be running right now. You open it back up. Spin up another agent. Fall asleep to the hum of tokens being consumed, and wake up reaching for the terminal before you reach for your coffee. This is token anxiety, and I'm deep in it.

The term

Nikunj Kothari coined the phrase in a February 2026 essay describing scenes from San Francisco's tech circles. A friend leaving a party at 9:30 on a Saturday, not because he was tired, but because he wanted to get back to his agents. Dinner conversations shifting from "what are you building?" to "how many agents do you have running?" The flex isn't what you've shipped anymore. It's what's working while you're sitting there not working. The concept resonated immediately. Bloomberg ran a piece on the "Great Productivity Panic of 2026," describing how AI coding agents have kicked off a high-pressure race to build at any cost. The term spread across dev communities, LinkedIn posts, and podcasts within weeks. It clearly struck a nerve because so many people recognized the feeling.

What it actually is

Token anxiety is the persistent feeling that if your AI agents aren't running, you're falling behind. It's the compulsion to keep compute occupied at all times, treating idle cycles as wasted opportunity. It blends the scarcity mindset around API costs with a deeper, more existential fear: that someone else's agents are shipping while yours are sleeping. There are a few distinct flavors of it:

The psychology underneath

What makes token anxiety stick is that it borrows from well-understood psychological patterns. Loss aversion makes idle compute feel like something actively being lost, not just unused. The sunk cost of a running subscription creates pressure to extract maximum value from every minute. And the variable reward loop of checking what your agents produced overnight mirrors the dopamine mechanics of slot machines. Jae Kaplan made this connection explicitly, arguing that coding agents trigger gambling instincts with slot machine-like behavior. You assign a task, walk away, and come back to see if you "won," if the output is good enough to ship. Sometimes it is. Sometimes it needs rework. But you keep pulling the lever because the next run might be the one. There's also something deeper happening with how we talk about these tools. People describe models the way sommeliers describe wine: this one has better taste, that one hallucinates with more confidence. They talk about harnesses and reins like they're controlling horses. We've started borrowing the language of animal husbandry for something none of us actually understand yet.

I'm in this loop

I'm not writing this from the outside. I'm writing it from inside the loop. So much productivity but no output. No product. Nothing worth showing. Forever running an endless race with myself. I spin up agents, I ship code, I optimize workflows, and at the end of the week I struggle to point at something concrete that matters. The activity feels like progress. The metrics say I'm busy. But busy isn't the same as building something real. The shift from human labor to machine labor has done something strange to my sense of accomplishment. When I wrote code by hand, finishing something felt earned. Now I review what agents produced and either approve or iterate, and the satisfaction is thinner. The work is faster but the meaning hasn't kept up. Kothari captured this perfectly: "Reading a novel feels indulgent now. Watching a movie without a laptop open feels wasteful." That voice in your head that says something could be running right now just doesn't shut off. And the worst part is, you can't even argue with it. It's technically right. Something could be running.

Why it's happening now

Three things converged in late 2025 and early 2026 to make this feel inescapable: Agents got good enough to run unsupervised. Tools like Claude Code and OpenAI Codex crossed a threshold where you could assign meaningful tasks and walk away. Overnight processing became real. That turned "should I be working?" into "should my agents be working?", and the answer to the second question is always yes. The capability curve steepened. Rapid model improvements mean that any workflow you set up this month might be obsolete next month. This creates a compounding urgency to stay current, to always be experimenting, to never let your setup sit still for too long. The social layer amplified it. Developer Twitter, now full of terminal screenshots and shipping receipts on Sunday mornings, created a visibility layer around productivity that didn't exist before. "What'd you ship this weekend?" replaced "what'd you do this weekend?" The performative pressure compounds the internal pressure.

The way out (maybe)

I don't have a clean answer for this. If I did, I probably wouldn't be writing about it at 2 AM. But I think the first step is recognizing that token anxiety is a real phenomenon with real psychological roots, not just a quirky tech culture observation. It's loss aversion meeting variable reward schedules meeting social comparison, all wrapped in a tool that's genuinely useful enough to justify the compulsion. The second step might be redefining what "productive" means. Occupancy isn't output. Running agents isn't shipping. And shipping isn't the same as building something that matters. The hardest discipline in an age of infinite compute might be learning when to let the machines sit idle, and being okay with the silence. Kothari's essay ends with a line that keeps coming back to me: "I still take aimless walks. The agents come with me now." The question is whether we can learn to leave them behind.

References