holding has no shape
i spent last tuesday doing nothing and came home knackered.
well, not quite nothing. i had six agents going at once, with one claude rewriting a migration while codex chased a flaky test through the integration suite, a couple more churning away in tabs i'd half forgotten about and one more running on my phone while i made a cuppa. i barely wrote any code myself that day. mostly i approved things and sent things back, read a great many diffs and said yes or no or told one of them to try again but leave the auth layer alone this time. by six in the evening i was completely fried, though not in the clean way that a long day on a hard problem leaves you. this was something worse and a good deal harder to explain.
that was the part that bugged me, because the tiredness didn't match the day at all. when i spend eight hours deep in one nasty bug i come out empty but fine, scrubbed clean in a way i actually like. this felt like the opposite, a scattered sort of buzzing that was still going at midnight, as though i'd spent nine hours holding a stack of things just off the ground and nobody had ever let me put them down.
nobody really warns you about this part of it. going from writing code to running agents looks like a promotion, but it isn't one, it's simply a different job. the work i do now has almost nothing in common with what i was doing two years ago and it taxes an entirely different part of the brain.
there's a paper from 1983 by lisanne bainbridge called "ironies of automation" she was writing about power plants and control rooms, nothing remotely to do with any of this and yet she caught something that maps almost too neatly onto it. when you automate the easy parts of a job, what's left for the person is the hard part, the part that only ever turns up when something has already gone wrong. you strip out everything routine and leave behind the stuff that demands judgement at exactly the moment there's no time for it. her point was that better automation makes the human's job worse rather than better, because the person ends up on permanent standby for the cases the machine can't handle and standby turns out to be far more draining than the work itself. i could never quite explain why that was until i read her.
that is more or less the day i had on tuesday. i sat there watching mostly competent systems do mostly correct things, waiting to catch the one that wasn't. and they do break, just rarely enough that you can never fully step away and often enough that you can never fully stop watching. at one point an agent decided the cleanest fix was to change a function signature that around forty other things depended on and it proposed this to me as though it were the obvious move, completely sure of itself and completely wrong. catching that took more out of me than writing the original function ever would have.
i don't want this to tip over into a rant, because i genuinely don't want them gone. they really have made me faster at things i used to be slow at. what i'm trying to point at is the price of that speed, which has a habit of showing up somewhere none of the dashboards are ever looking.
the mechanism is fairly dumb once you lay it out. every new agent feels free, because you're already paying for it and it's sitting there idle, so you hand it the annoying little task you've been avoiding and then the one after that. each decision to spin up one more is sensible on its own and costs you next to nothing. the trouble is that they don't come back one at a time and in order, they all surface at once needing a decision from you and the cost was never in starting them so much as in holding them all in your head at the same time. you become the working memory for six parallel threads and human working memory was barely built for two.
that is the part i keep returning to, the reason it drains you rather than simply tiring you out. effort at least has a shape to it, in that you push against a problem and it pushes back and eventually one of you gives way. holding has no shape at all, only a steady low load that never resolves and your attention gets stretched so thin across the threads that switching between them starts to leave a residue. you come back to the migration agent and find that the constraint you set ten minutes ago has vanished, because in those ten minutes you were off being three other places at once.
the bottleneck used to be how fast i could write code. for a while after that it was how fast i could steer the things writing the code. now it is something smaller and more uncomfortable than either, namely how many of them i can hold in my head at once before one of them slips. the limit has moved off the machine and onto me and there is no subscription tier that fixes that particular one.
so what do you actually do, on a tuesday like that, when this is simply the job now and the work still has to ship regardless. i don't have a tidy answer, but i do have a few things that have helped and the honest version is that all of them are about admitting the limit rather than trying to beat it.
the first is just a number. i now cap it at two or three agents rather than six, because the moment i'm past three i can feel the quality of my attention fall away and a fourth agent producing work i'm too fried to review properly. past a certain point more parallelism stops buying speed and starts buying mistakes that i'll only find later, at a worse time. how many you can actually run turns out to be a fact about your attention rather than your wallet and pretending otherwise is how you burn yourself out without noticing.
the second is that i stopped watching them work, which sounds like a trivial change and really wasn't. there is a strong pull to monitor every output as it streams in, to sit there nodding along and that constant real time vigilance is the single most depleting part of the whole business. so instead i batch it now, letting three agents run while i go and do something else entirely, then coming back to review the lot in one focused pass. you trade the illusion of control for some recovery and the work is honestly no worse for it.
the third one i'm more sure of than the other two. every single day i protect a block of time where nothing at all is running, no agents or review queues, instead me and one problem and the slow good work of thinking it through by hand. part of that is for my own sanity, but it's also because the thing that lets me smell a bad function signature change coming, which is the ability to hold a whole problem in your head at once, is a muscle and supervising agents all day trains an entirely different one. if i let that deeper muscle go soft, i lose the exact judgement that makes me worth keeping in the loop in the first place. you can't stay sharp as a reviewer once you've stopped being a builder and i've become fairly convinced the two either feed each other or starve together.
there is, of course, a version of the future in which the agents simply get good enough that they stop needing me to catch their mistakes at all and plenty of clever people are betting on exactly that. they may well be right. but bainbridge's irony has held up for forty years across every kind of automation we've tried and the pattern she described was never really about the machines getting better, it was about what happens to the human when they do. so far, every round of automating the easy work has only made the residue that lands on the person harder and lonelier and more exhausting and getting better at automation has sharpened that rather than softened it. i've yet to see it run the other way.
i'm not arguing for fewer agents in any permanent sense and i'm certainly not giving back the speed. i've simply stopped treating my own attention as the free variable in the equation, the thing you can draw down endlessly while everything else scales up around it. it is the scarcest resource in the whole setup and it's where the real bottleneck has been hiding the entire time.
most days i still feel like the bottleneck and lately i've started to think that's the system working as intended rather than failing. the machines are built to run wide and i am built to go deep and the whole arrangement only holds together because there is still one person in the middle who can actually think.