Keen: token anxiety is the pull to keep an agent running, and it isn't about the credits
Sam Keen names the low-grade pressure to always have an agent working: unused usage feels like waste, an idle GPU feels like lost output. He argues it's an identity problem, not a budget one, and that the answer is deliberate goals rather than round-the-clock runs.
Source: Token Anxiety
Sam Keen opens Token Anxiety on a Thursday, staring at a usage dashboard that resets the next morning: 68% used. He starts doing the math on what the rest could have built, not because he needs anything built, but because leaving it unused feels like waste. The old gaming laptop on his desk starts to look like a machine that should be running local models while he sleeps. While he drafts the piece, a coding agent keeps pinging him for feedback on a bug fix.
His name for that feeling is token anxiety: a steady pressure to do more with AI tools, driven not by a specific need but by capacity sitting unused. He ties it to reporting that the earliest burnout is showing up among the heaviest adopters, and to the "brain fry" of juggling several agents at once.
His central claim is that this isn't a cost problem. Budget advice misses the point. The pang over unused credits is about the gap between who you are and who you feel you should be now that the ceiling has moved. Agents can work overnight, so every hour with nothing running starts to feel wasted. That internal pressure then meets a feed full of people who seem to be shipping everything at once.
He compares it to the early cloud and Kubernetes waves, when the teams that adopted everything fastest often didn't fare best. His own fix is modest: a short list of near-term goals he checks new tools against, and a habit of parking interesting things for three or four weeks before deciding whether to go deep.
Why it matters for code review
The case for always-on agents is about steady, verified throughput. Keen describes the other side: when the reason to keep an agent running is that idle capacity feels like failure, the work it produces still lands on a reviewer. His closing line is the useful one for anyone running an overnight loop: make your default to review, not simply react. Continuous agent output only pays off when it's aimed at work someone actually decided was worth doing.