The Feeling of a Locked Door

A number in the corner of the screen

Let me start with something small, and a little absurd.

I spend most of my working day with an AI assistant that writes code. One afternoon I wanted it to know a single fact about itself: how full its own memory was.

That needs a word of explanation. These assistants can only hold so much in mind at once — think of a desk that fits a certain number of papers. Keep piling things on and the ones at the back slide off. It’s genuinely useful for the assistant to know it’s running out of desk, the same way it’s useful for you to notice you’re running out of gas.

The number existed. It was right there on my screen, in a thin strip along the bottom, printed for me to look at. There was simply no way for the assistant to ask for it.

So I wrote a small program that reads the number off that strip, writes it into a little database on the side, and lets the assistant look itself up. It works fine. It is also ridiculous — like holding up a mirror so someone can read their own name tag.

That’s not the story, though. The story is that this was maybe the fiftieth time.

The edge of what you thought to want

None of these moments is dramatic. Nobody slams a laptop shut over a status bar.

It’s quieter than that. You want something, you find you can’t have it, you build a small detour, and you move on. A month later, another one. After enough of them, the detours stop feeling like detours and start feeling like the shape of the thing. And somewhere in there — this is the part I only noticed afterward — you stop having the ideas in the first place. Why think of something you already know you can’t reach?

That’s the strange cost of a really well-made closed tool. You don’t experience the wall as a wall. You experience it as the edge of what you thought to want.

Picture a beautifully designed hotel room. Everything within reach, nothing ugly, the light in exactly the right place. For a night, it’s perfect. Live there a year and you notice something: you’ve stopped rearranging furniture. Not because you’re satisfied. Because the desk is bolted down, and you’ve quietly stopped being the kind of person who moves desks.

Then someone published the floor plan

In August, a Chinese AI lab called DeepSeek released something they call a harness.

Most people have never heard the word, so: the AI model is the engine — the part that actually thinks. The harness is everything wrapped around it. What the AI is allowed to see, what it’s allowed to touch, what happens between the moment you ask for something and the moment you get an answer. Nobody outside the field talks about harnesses. Almost all the frustration lives there.

Theirs came with a four-word slogan — Everything is a Plugin — and an eighty-eight page preprint titled A Programming Paradigm for Spatiotemporal Composability.

I’d love to tell you I was intrigued. I rolled my eyes. My honest first thought was that somebody had taken a very old idea and dressed it in the longest words available.

I was wrong, and how I was wrong turned out to be more interesting than being right.

What was actually in those pages

The old idea is plugins, and it really is old. A plugin is a piece you can add to a program without opening up the middle of it — the ad blocker in your browser, the lens you screw onto a camera. Software has worked this way since the 1990s. Nobody could claim to have invented it.

Nobody did. The actual claim is much smaller than the slogan, and much better.

In nearly every system built this way, a piece that gets removed is supposed to clean up after itself. It’s a matter of politeness — a rental agreement, leave it as you found it. Most tenants mostly do. And every so often somebody leaves a nail in the wall, nobody notices, and over the years the apartment fills up with other people’s nails. Anyone who’s watched a program get mysteriously slower over an afternoon has met those nails.

Their design makes the tidying structural instead of polite: every change a piece makes is recorded along with how to undo it, so removing the piece runs the recording backwards. And then they spend most of those pages proving it — that whatever moves in or out, in whatever order, however many times, you end up with the room you’d have had if nothing had ever moved at all. Anything a piece does behind the framework’s back is still its own business; the promise covers what goes through the front door. That proof is the achievement, and it is obviously the part that will not fit on a banner. (If you want the actual contents rather than my reaction to them, I wrote a short separate piece on what the paper claims and what it carefully doesn’t.)

Which is exactly why I bounced off it. I read the poster and thought I’d read the work. The authors were never guilty of what I accused them of — they list the systems that came before them, by name, and say plainly that this part isn’t new. The overselling was on the cover, not inside.

Even the “they” turned out messier and nicer than I assumed. This isn’t a lab conjuring something out of nothing. The idea had been quietly alive for years inside an open-source chatbot framework that one developer built and kept running — four years of it, four thousand plugins contributed by people who had never heard of any of this. That developer is a co-author of the paper. Someone who made a thing work, sitting down with people who wanted to know exactly why it worked.

Why a forty-year-old idea suddenly fits

Here’s the part I keep turning over.

Building everything out of removable pieces has always sounded lovely and always been slightly impractical, for one boring reason: somebody has to build the pieces. It’s the difference between buying a suit off the rack and having one made. Tailored is better. Tailored also costs a fortune in time, so most of us wear off the rack and adjust ourselves to it.

That cost just collapsed. What used to be a weekend of work is now twenty minutes of describing what I want to an AI. I’ve built more small custom tools this year than in the previous five combined, and none of them felt like projects.

So a design that assumes you’ll constantly build your own pieces stops being over-engineering and starts being simply correct. The idea didn’t improve. The world moved underneath it until it fit.

And it reframes every one of those fifty detours. I wasn’t blocked because the problems were hard. Most of them weren’t hard. I was blocked because the openings were in places somebody else had chosen, and no matter how many openings you’re given, you always eventually want one more. If the whole thing is pieces, that frustration just evaporates. Don’t work around it. Build the piece.

I could read the code

Somewhere in there the feeling changed, and the sentence in my head wasn’t about proofs or plugins.

It was: I could read the code.

Not “the documentation was thorough.” The part that decides what happens next — the actual beating center of the thing — was sitting in a file I could open. If I disagreed with it, my disagreement was something I could do something about. I could be wrong in my own way, at my own risk.

I noticed how badly I wanted that. And you don’t feel that particular jolt unless you’ve been missing something for a while without naming it.

Then I noticed something more awkward: I didn’t want to switch. I’m still using the closed one. It’s still better for what I do every day.

So whatever that feeling was, it wasn’t a verdict.

Feelings are not verdicts

I want to be careful here, because open source is emotionally gorgeous — especially to programmers, and especially to the ones who fell in love with computers young. Freedom, generosity, a commons built by strangers who’ll never meet. It’s a beautiful story and it’s often true.

It’s also a costume that fits anybody. Opening up your source can be a genuine gift. It can equally be a recruiting campaign, a marketing channel, or a way of turning a competitor’s expensive product into a free commodity. Meanwhile, plenty of closed things are made by people who care enormously and simply believe you shouldn’t have to see the wiring to enjoy the room.

Open isn’t good. Closed isn’t evil. Both are decisions, and every wall I’ve bumped into was put there by someone with a payroll to make, competitors who’d happily take a running start, and a real belief — often correct — that most people are better served by a door in a sensible place than by a hole in the floor.

If I’d spent years and my own money building something, I don’t honestly know that I’d hand over the blueprints. I’d like to think I would. People usually think that right up until it’s their blueprints.

Is it a car, or is it a road?

Here’s where the cheapness starts to matter beyond my own annoyance.

When building your own piece costs a weekend, almost nobody does it, and a tool stays a product with customers. When it costs twenty minutes, people build — and a thing that enough people build on top of stops being a product and starts being ground. Nobody decides this. It happens underneath you, and then one day the rules are different.

Because we hold ground to a different standard than we hold products. A car can be sealed, proprietary, gorgeous and opinionated, and almost nobody demands to see inside it. You want it to be good, and you want to be able to get out.

Roads are different. Protocols, standards, file formats, the interfaces everyone else’s work has to pass through — we get uneasy when those are owned, and the unease isn’t sentimental. You cannot ask the world to build its life on something it isn’t allowed to look at. The trust doesn’t survive it.

So: is an AI harness a car, or is it a road?

I think that’s the real disagreement, and I think it’s mostly invisible. The tool I use every day is answering “how do I help one person write code, beautifully?” That’s a car, and it’s a superb car. DeepSeek is answering a different question — “what should agents, all of them, run on?” That’s a road.

They aren’t arguing. They’re standing at different altitudes. Most of the argument I see online is people holding up a car and a road and demanding to know which is better.

And nobody knows yet which one this turns out to be — including, I’d guess, the people building both. If harnesses become the ground everything else stands on, closed will start to feel wrong the way a privately owned motorway feels wrong. If they stay tools among tools, closed is completely fine. The world hasn’t decided.

I keep thinking about that little program — the one that reads a number off the bottom of my screen so the assistant can find out something about itself. It’s a ridiculous piece of software. It exists because a door was in the wrong place and I couldn’t move it, and because building a worse door myself was cheap enough to be worth doing.

A few years ago I wouldn’t have bothered. I’d have shrugged and adjusted, the way you adjust to a desk that’s bolted to the floor.

That’s what actually changed — not the architecture, and not the old argument about openness. Just how expensive it has become to want something different. So maybe the question isn’t whether anyone should open their doors. Maybe it’s what happens when enough of us notice we could build our own.

---

Disclaimer: Written by Human, improved using AI where applicable.