Why a good assistant goes blind to you

I build an assistant whose best trick is taking over the small things you do on repeat. You do something often enough, it notices the pattern, offers to handle it from now on, and once you agree it happens on its own. I’ve been calling that habit hijacking. To pull it off, the assistant has to learn you from your behavior, from what you do day to day, because a habit only shows up as a repetition. The better the thing works, the less of you it can see.

Working has a built-in catch. The assistant spots that every Monday morning you stitch the same few documents into a summary, and offers to have it waiting before you sit down. You say yes. From then on, you don’t do it. The habit it just learned drops out of your behavior the moment it takes over. It performs the routine out of existence. Every habit it successfully hijacks is one it can no longer watch you have.

So the system is quietly eating the thing it runs on. It learns from your repetitions and then removes them, one by one, in order of how repetitive they were, because the most repetitive ones are the easiest to catch and the most worth taking. What’s left in plain view is everything it hasn’t automated yet: the one-offs, the irregular, the brand-new. Over time the part of you that’s legible to your own assistant shrinks to whatever you haven’t handed over. It knows you best right before it helps you, and a little less after.

This flips the way people usually think about these assistants. The assumption is that the more it does for you, the more it understands you, that capability and closeness grow together. I think it’s closer to the opposite. An assistant that has absorbed your whole routine is left watching a person who, from its side, barely seems to do anything the same way twice. You’ve become mostly noise to it, and that is what its success looks like from the inside. The signal it needs is the signal its success destroys.

There’s a smaller, practical version of the same trap. An assistant like this is only as current as the behavior it can still see, so the longer it runs, the more it’s working from a portrait of you assembled mostly before it started helping. Your habits keep drifting, but a lot of that drift happens inside the actions it has already taken off your plate, where it can no longer watch. Left alone, it slowly serves the person you used to be.

Once you see this, it changes how you’d build the thing. The assistant can’t lean on watching you to know what it’s already taken over, because that evidence is gone. It has to keep its own record of the habits it absorbed, and read the silence where a habit used to be as a signal in its own right. An empty Monday morning carries information: the trace of a routine that’s now handled out of sight. The thing has to learn to read its own footprints alongside yours.

What I find strangest is the version of this that isn’t about software at all. Carried to its limit, the most stable, repeated parts of your life become the parts that are handled, and so the parts you stop seeing. What stays visible, to your tools and maybe to yourself, is the residue: the things you do once, and the parts of your life that haven’t settled into a routine yet. The better everything runs, the more you’re defined by whatever you haven’t automated. I don’t know yet whether that’s freeing or a little sad. Probably both.