A House Built for Robots

I had a funny thought the other day. Everyone keeps asking what software engineers are going to do with all the time AI is supposedly giving back to them. All those productivity gains — what do we spend them on? More features? More products? Long lunches? Touching grass?

I think I figured out the answer, and it’s a little cursed: we spend that time learning about and managing the tools that gave us the time in the first place. The savings get reinvested directly into the thing that produced the savings. It’s a closed loop. You save four hours, and you spend four hours figuring out why your agent decided to refactor the auth system unprompted.

I should say up front that I’m not throwing stones from outside the house. I’ve spent the last year writing posts with titles like The Session Is the Unit and The Layer Above the Model, and the whole argument of those is that the harness around the model is the thing worth investing in — the part that compounds while the models churn underneath it. I have a workflow with phases and handoffs. I have opinions. I am exactly the person I’m about to describe. Which is the reason I can describe him so accurately.

AI agents are not the first autonomous, always-on, self-directing intelligence we invited into our lives to do the boring work for us. That distinction belongs to the robot vacuum.

Meet the robot

You buy a robot vacuum. The pitch is beautiful: the robot does the boring thing, and you get your Saturday mornings back. Then reality sets in.

The robot, it turns out, has needs. You need to keep it updated. You need to use its proprietary app. You need to learn how it sees the room — which is to say, badly. You need to set up its navigation paths. You need to put down little boundary strips, which are just tiny walls you’ve built inside your own house to teach the robot where it’s not allowed to go. You need to buy furniture risers so it can fit underneath the couch. The ottoman keeps getting headbutted, so now you need to move the ottoman.

That’s the trade nobody mentions at purchase. It’s a slippery slope, and it works because every step is individually reasonable and every step is justified by the one before it. You bought a vacuum, and a month later you’re paying someone to mount the TV on the wall and fish the cables inside it, because cords on the floor confuse the robot — and you could not honestly say that was your idea.

That’s the part that sneaks up on you. At some point you’re not arranging your living room for you anymore. You’re arranging it for the robot. A guest walks in and says “nice place,” and you can’t quite explain that none of this layout is for human beings. It’s for the robot.

The same thing, with a keyboard

The AI-assisted development version is this exact story, and the furniture is your codebase.

You start writing smaller files, because the agent gets lost in big ones. You add obsessive docstrings and comments — not for the next human, for the model’s context window. You name things more explicitly than you naturally would. You break functions down past the point of your own taste, because it parses better that way. You write a whole CLAUDE.md or a rules file, which is, functionally, a map of the house taped to the wall so the robot doesn’t get stuck under the couch.

None of these are bad on their own. Some of them are even good hygiene. But notice the direction the pressure runs. The architecture starts bending toward “what does the model handle well” instead of “what’s elegant” or even “what do I find readable.” You’ve started rearranging the furniture for the robot, and you didn’t decide to do it on any particular Tuesday. It just happened.

The irony that eats itself

Here’s what makes it genuinely funny instead of just annoying: the productivity boost is exactly what creates the pressure that eats the productivity boost.

If these tools only made you five percent faster, nobody would panic. You’d ignore them, or you’d adopt them at your leisure. It’s because they’re a step change that there’s an arms race — and the arms race is the thing that sends you down the rabbit hole at eleven at night. What IDE? Cursor, or the agent in the terminal? What does your rules file look like? What skills are your agents loading? How are you orchestrating a swarm of them? What do your loops look like, your cron jobs, the handoffs from one agent to the next?

Back to the vacuum: it got so good that now everyone on the street has one. There’s a neighborhood newsletter for advanced robot-vacuum technique, somebody’s posted their multi-floor routine — the upstairs robot coordinating handoffs with the downstairs robot — and now you feel like a Luddite for owning only one vacuum and not having rearranged the whole house yet. Clean floors stopped being impressive a while ago. What’s impressive now is the orchestration.

And almost none of it is settled. There’s no right answer to most of these questions, because the whole landscape rerolls every few months. So you’re not building durable expertise — you’re sprinting to master a setup that has a real chance of being deprecated by the time you’ve got it dialed in. You optimized your furniture for a navigation algorithm they’re about to retire.

We have always had this itch

Step back from the arms race for a second and you’ll realize that we were never going to bank that time. Saved time doesn’t accumulate in an account somewhere — it gets filled, immediately, by whatever’s nearest and most absorbing. We always find something to pour the hours into. This time it just happens to be the thing that gave us the hours in the first place.

The thing is, this didn’t arrive with AI. The AI harness is just the new dotfiles.

This is the same person who spent three weeks in 2016 perfecting a Vim config. Tiling window manager. Terminal prompt showing the git branch in exactly the right shade of green. Mechanical keyboard with hand-lubed switches. Beautiful, every bit of it. What are they shipping? Couldn’t tell you.

The tell was always the same, and it’s worth saying plainly because it’s the actual point of all this: the setup is the work, emotionally. Configuring is comfortable. It has clean, well-defined progress. Every time something snaps into place you get a little hit of satisfaction. And — this is the real draw — nobody can reject it. A customer can hate your product. A market can shrug at the thing you spent a year on. Your terminal prompt cannot hate you back. So if you’re at all afraid of the verdict, the environment is the perfect place to hide, because fiddling with it feels like work, it photographs like work, and it never makes you find out whether anyone wants what you’re making.

AI just supercharged the hiding spot. “I’m building out my agentic workflow” sounds so much more legitimate than “I reorganized my plugins again.” But it can be the exact same avoidance in a better jacket. The orchestration diagrams are gorgeous. The robot ballet is mesmerizing. And meanwhile the floor — the product, the thing someone actually pays for — is off-screen, and nobody’s looked at it in a while.

The unglamorous win

So who actually wins?

Mostly the people who pick a setup that’s good enough and then deliberately stop fiddling. Which is the single move the entire discourse is built to make you feel insane for doing.

You can have the most elegant AI harness on the street and have shipped nothing. A hypnotic swarm of agents moving in perfect coordination across a repo that does not contain a product anyone wants. It’s possible to admire how beautifully the robot moves and never once check whether the floor is clean.

The people who win have a higher tolerance for the ugly part. Talking to a confused customer. Shipping something a little embarrassing. Leaving the harness alone today, even though they could make it nicer, because nicer isn’t the job.

Get a vacuum. Let it do a decent-enough job. Resist the urge to rebuild your house around it.

You got the hours back. Try to save a few for yourself.