Andrej Karpathy told the No Priors podcast he's been in a "state of AI psychosis" since December. He now spends 16 hours a day issuing commands to agent swarms, with his ratio of hand-written to AI-delegated code moving from 80/20 to 0/100 in a matter of weeks. Armin Ronacher, Flask and Rye creator, wrote in January: "Many of us got hit by the agent coding addiction. It feels good, we barely sleep, we build amazing things." Steve Yegge wrote that agentic coding is "genuinely addictive," that he suddenly crashes and falls asleep mid-session.
These are not struggling developers. They are among the most technically sophisticated people working with AI right now, and they are describing something that sounds less like a productivity upgrade and more like a compulsion.
Why This Loop Is Different
The obvious framing is "developers love building, always have, nothing new here." But that misses the specific mechanism at work.
Flow state — the psychological state that makes deep coding sessions sustainable and rewarding — has a recognizable signature: it's predictable. You sit down, you build context, the work has internal logic that you learn and extend. The feedback is incremental and tied to your own comprehension. You can feel it working. You are the agent.
Agentic coding has a different structure. You prompt. You wait. Sometimes the output is mediocre or wrong. Sometimes it's exactly what you needed plus three things you didn't know you needed. The quality is non-deterministic — not random, but variable enough that you can't predict which prompt will produce the "holy crap" moment.
This is the slot machine schedule. Behavioral psychologists call it partial reinforcement: intermittent rewards at variable ratios produce the most persistent behavior and the hardest-to-extinguish response. It's why slot machines are more addictive than games with fixed payoffs. The uncertainty doesn't discourage people — it locks them in. Maybe the next pull will be the one.
With agents, the next pull is another prompt. The next feature, the next refactor, the next attempt at something you couldn't do before. Ronacher described the addictive hit precisely: it feels good, you barely sleep, you build amazing things. But you don't always build amazing things. Sometimes the agent loops or hallucinates or misses the constraint you forgot to specify. The variability is the mechanism.
The People It Hits Hardest
The developers most caught in this loop are not the casual users. They're the power users — the ones who have already hit the jackpot and know it's achievable.
This is another slot machine feature: the machine is most seductive for people who have won before. When you have experienced an agentic session where two hours produced what would have taken two weeks, you know the ceiling is real. That knowledge pulls you back in. Every mediocre session carries the implicit question of whether the next one will be the extraordinary one.
Axios spoke with Claude Code and agentic tool users in early April on exactly this pattern. One developer described needing a doctor's help to turn off cerebral activity that wouldn't stop at night — running through prompts mentally, strategizing next sessions, unable to let the context go. Several described waking up to check agent output. The framing was exactly right: these tools operate like slot machines, and the people who can least afford the cognitive cost are the ones most invested in pulling the lever.
What the Cost Actually Looks Like
BCG's cognitive load research from March tracked 1,488 full-time workers on the effects of AI tool use. When AI-related work required high oversight — checking output, managing parallel agents, validating code — workers expended 14% more mental effort, with 12% greater fatigue and 39% more major errors. Those numbers are from daytime work, with presumably reasonable sleep behind them. Extend the session into night — baseline sleep degraded, cognitive load accumulated all day, dopamine loop running — and the error rate goes where you'd expect.
What most developers running late agentic sessions don't track: whether the code committed at midnight is the code that ships. In developer time-tracking data, late-night sessions show meaningfully higher rates of next-day reverts. The session felt extraordinary. The morning-fresh read of the diff told a different story.
The out-of-hours commit numbers support this pattern at scale. Researchers tracking AI-era developer behavior found a 19.6% rise in out-of-hour commits since widespread agent adoption. More code getting written at 11pm is not the same as more good code getting written at 11pm. It means more of the slot machine cycle running outside the hours when your judgment is reliable.
What to Watch For
The goal is not to avoid agents. The output when it works is real, and Karpathy's 0/100 ratio is producing genuine things. The point is that "genuinely addictive" describes a reinforcement schedule, not a productivity endorsement.
A few signals worth tracking explicitly rather than running on feel.
Your late-night session rate. What fraction of your total coding hours happen after 9pm or before 7am? If that number has increased since you started running agents, and especially if it's trending upward, you're seeing the schedule effect in your own data.
The revert rate on off-hours code. How often does work committed in a late session get reverted or substantially reworked by morning? Anything above 20% is worth treating as signal. The session felt productive. The morning review is the more reliable judge.
Sleep quality against agent usage. BCG's fatigue effects accumulated across days, not just sessions. If you're wearing a heart rate monitor or tracking HRV — Whoop, Apple Watch, Garmin — look at your recovery scores on days following heavy agent sessions versus conventional deep work days. The relationship is usually not subtle once you look for it.
The developers navigating this well are not quitting agents. They're setting explicit cutoffs, tracking the metrics that don't flatter the session, and treating the itch to open Claude at midnight as the signal it is: the loop talking, not the work calling.
You cannot build sustainable output on slot machine cycles. The jackpot feeling is real. So is the house edge.