# The high wears off
Table of Contents
I used to not open my laptop after 9 PM. Now I do, most nights, to wrap up a Claude Code session that got cut off during the day. Fifteen minutes, tops. It’s never fifteen minutes. I go to bed too late and wake up tired. I know the pattern. I do it anyway.
The last post covered how individual speed gains collapse at the team level. This one is about what happens to you when you can run AI agents in parallel all day and the only constraint on how long you keep going is your willingness to stop.
More parallel, more drained
You’re not doing the same cognitive work faster. You’re doing a different kind of work. Two parallel agents means two problem spaces held in working memory simultaneously, each demanding real-time review and judgment. Three means you’re not really holding any of them properly. You’re skimming, trusting, letting things through. The work shifted from building to supervising, and supervision at the speed of code generation drains you in a way that building never did.
Steve Yegge put it bluntly: intense AI-assisted sessions produce a dopamine-adrenaline high, followed by hard crashes. He ran 20+ Claude instances and experienced sudden “nap attacks” mid-afternoon. His colleagues are apparently considering nap pods. I don’t run twenty instances, but I recognize the pattern. A sense of almost limitless capability during the session, then a crash that feels completely out of proportion to the effort I thought I put in.
Simon Willison described the same: firing up four agents in parallel and being wiped out by 11 AM. He’s talked to people losing sleep over this, staying up an extra half hour to kick off more agent runs, waking up before their alarm to check results. The practical ceiling for most people seems to land around three to four focused threads before judgment quality degrades. On a tired evening, it’s closer to one.
Last week I wrote about teams rubber-stamping each other’s AI output. Turns out you do it to yourself too. Three sessions deep into an evening, I stop questioning what the agent produces. Not a deliberate choice to trust it, just a lack of energy to do anything else. And you don’t notice, because the dopamine loop tells you you’re still being productive.
So why don’t you just stop?
The slot machine
Every prompt is a lever pull. Sometimes you get clean, elegant code. More often you get something 80% right: slightly off variable names, incomplete error handling, an edge case quietly ignored. Close enough to feel fixable. The fix is always one prompt away.
I’ve spent entire evenings in this loop. It usually starts honestly. The session hit the token limit at 6, the bug was nearly cracked, I’ll just finish what I started. One more prompt. One more refinement. An hour later the evening isn’t about shipping anymore. It’s about the pull of the lever.
The uncomfortable part is that the people most susceptible to this aren’t the weak engineers. They’re the best ones. Using coding agents well takes every inch of your experience as an engineer, and it is mentally exhausting. The people with the most skill and drive are the ones who keep pulling the lever when a saner person would close the laptop.
The limit was the feature
Remember a few months ago, when the $20 plan ran out of tokens mid-afternoon and you had to go do something else? At the time it felt like friction. In hindsight, those breaks were where the actual thinking happened. You’d step away from the terminal, and somewhere between the coffee machine and the whiteboard you’d realize the approach was wrong, or that the feature didn’t need to exist at all. That’s the part AI can’t do for you, and it only happens when you stop prompting long enough to reflect. The rate limit forced that reflection. Without it, you just keep going.
But it’s hard to ignore that the people most AI-pilled are working harder than they’ve ever worked. Meta built an internal leaderboard ranking 85,000 employees by token consumption, turning usage into a performance signal. The industry term is “tokenmaxxing.” That’s where we’re headed when nobody builds in a reason to stop.
The problem isn’t that AI makes you less productive. It’s that it makes you productive in a way that’s unsustainable, and you mistake the high for health.
Protecting the scarce resource
If cognitive endurance is the actual scarce resource in the AI era, and everything I’ve experienced confirms it, then the countermeasures aren’t productivity hacks. They’re about reimposing the constraints that unlimited plans removed.
Not every parallel run is destructive. Independent test generation, dependency upgrades, doc passes, anything embarrassingly parallel and judgment-light is what these agents are for. The drain comes from supervising parallel thinking: two architectures, two debugging trees, two product decisions held in working memory at once. That’s the work to limit, not the boring fan-out.
The engineers who burn brightest with AI aren’t the ones who’ll last. The ones who figure out their limits, who can still think clearly at dinner, will still be building great software five years from now.