There’s this concept in Computer Science called thrashing. The tl;dr is that thrashing can happen when programs running on a computer are actively using more memory than can fit in RAM. The OS will keep moving data between memory and disk, making room for something by removing something else (that it will need again shortly). And, at some point, so much time goes into moving things around that the programs barely make any real progress.
Boy, oh boy, isn’t that every single one of us right now using AI? One agent is working, and you, almost by reflex, will start something else. Kind of like doomscrolling: you don’t even realize you’re doing it.
And by the way, of course we’re doing it! These apps (Codex, Claude Code, Cursor, etc.) make it so easy to keep several things going at once. It feels silly to be sitting there watching the agent build your thing while you’re just passively watching.
But I think this exact behavior is what’s making a lot of us feel exhausted at the end of the day. Mostly because of that last part: multitasking.
Because when the agent comes back, you need to understand what it did before you can do anything useful with it. You open the diff 2 hours later and try to remember what you asked for… and right when you’re starting to make sense of it, another agent (with something completely different) now wants your attention 🤡
So you go there because it’s just a quick question (you think). But answering it requires remembering what that other agent was doing and why you chose that approach, etc. By the time you come back to the first one, you have to restart the whole process again.
You can spend a surprising amount of your day doing this! Even when the agents are doing a great job, you still need to get enough of each problem back into your head to know whether their output is making any sense…
I think we’re making this worse by saying yes to things we would’ve simply put on our priority queue (or just left alone) before. Because “I’ll just ask my agent” feels like such a small commitment — and easier than creating a ticket, right? — that we forget we’re also volunteering to understand, check, and finish whatever comes back.
(Btw, this reminds me of the Jevons paradox where using a resource more efficiently can lead to consuming more of it overall. Jevons was talking about coal, but I think there’s some similarity here where each task gets cheaper, so we take on more of them)
Don’t get me wrong, I still pretty much want to use AI, but I do think we need a better/more sustainable way. One OS approach when it comes to thrashing is to suspend some processes, so the others have enough memory to make progress. And maybe that’s what we should be doing as well?
To make it a bit more practical:
- Decide if it’s worth taking on now. Before asking an agent to do something, take a second to decide if you actually want to take that on right now or if it can wait. What else are you already working on? Is this important, or did it just become easy enough that you stopped asking that question?
- Remember that you still have work to do. Include your part when figuring out how long something will take (that one where you’re using your brain to verify AI. I mean, you’re still doing it, right?). Just remember that the review time needs to fit somewhere in your day too.
- Finish some of what you already started. If you already have a bunch of things waiting for you to review, go finish some of those before starting more. It’s tempting to treat “the agent is done” as “the task is done,” but it’s really, really not; there’s still work to be done here (and it’s yours).
I think that, at the end of the day, we might need to get comfortable with just letting an agent sit there with nothing else to do sometimes. Finish the task, make sure it works, and only then move to the next thing.
We don’t need to keep 10 agents busy all day every day just because we can; that feels like productivity but can leave us spending the whole day trying to catch up with work we asked for ourselves. And honestly, I’d like some of the time AI saves me to actually feel like time I got back.