In which the interface no longer matters, perhaps.
The Interface Was the Meter
The Interface Was the Meter
Why check my email in the morning? Why start the ritual? It has nothing to do with what's actually in my inbox: a lot of pitches to sell me something, a few problems not yet resolved, a headache, a nice note, an easy resolution to an issue I was worried might be a bigger problem. The message I chose not to answer is still there, and every non-urgent urgent thread is still getting endlessly, annoyingly pinged to my deep chagrin because it's not urgent and it's borderline important.
What nobody told me is that the inbox was never for me. It just works better if I think it is.
The world gets to fill your inbox at its discretion. You are almost entirely reactive to it. On a good day you zero it out before it starts to refill.
I check my email and fill a "seat." I log in and do something the email system can see: click around, archive, move some labels around, respond, compose. And the company on the other side (Google) sees all that happening somewhere, in some system, and assumes I will be back the following month and with some more math applied, the following year. I can't work without the interface, and working through Gmail is what Google is counting. Sort of like the meter on the side of a house.
When I joined my first startup, Design was already a strategic function. They held real power at the company. Screenshots became marketing. UX polish became how our tool was differentiated. It took me a bit to realize that companies hired designers because the interface was where our customers felt the level of competence that parted dollars from their wallet. It took me forever to draw that connection because I was more interested in everything underneath the interface: databases, APIs, permission models, audit trails, automated workflows. All those layers truly mattered, but they were largely invisible to the people the software was built for and around.
For us, the interface was where our customer's work was counted: who clicked what, did what, what actions signalled something back to us. And as we counted seats, the managers and team leads on the other side were counting heads. And so we had a common item by which to discuss pricing. Both sides could look at who needed the tool and understand, roughly, how much of their company's value was flowing through that particular system. Today, increasingly more and more "heads" flow through the systems, but it's not humans doing the clicking.
Demotion of the Screen
I have my system at work, which I mostly follow every business day: one by one, I open the same browser tabs in the same order, the same checks each time. Sometimes I feel like a guy in a nuclear plant or in the mission control center. CRM, check, report, check, reports, check, meeting prep, check, spreadsheet, check, email, borderline but check, slack, out of control but not launch blocking, etc., etc. Maybe not mission control, but like watching Ichiro warm up before getting ready to bat. Or going for a run every day.
Earlier in the LLM days, you could inspect the thinking, then everyone realized the thinking itself was valuable training data for everybody else, including the Chinese labs, and so the models mostly stopped showing it. And now I can barely ever understand how we got where we got together. Because it's usually on the hardest problems where I want to know the thinking, where I have the most to learn, but my single instruction can step in where I used to have to walk tab by tab. A system of systems decides where to go and what to call to give me what I need, how to reconcile it, and what to return, often running where I or anyone else would have to walk.
I used to verify. Now I'm a supervisor, which is what it's called when you're not always exactly sure about what you're looking for but definitely feel responsible if something goes wrong.
The sequence did something else; it made everything easily inspectable. The tabs are still there if I go looking for them, and the numbers are still where they used to be, but I don't have to always move through them in the same way or watch them change as I go and now I can actually get a lot more done and still do the stuff I was doing before. Which is an odd headspace to be in. Simultaneously thrilled and wowed by having effective superpowers compared to your prior self, but now not sure the "things you used to do carefully before" are still being done as carefully.
So what does this mean? Shrug. Don't know. But there's something in the fact the day stops beginning in apps, or they don't feel as important to "me" but to "my agents". The interface and the work are starting to come apart. And the seat is still there, but the work no longer passes through it in a way that can be easily counted.
Rerouting
We're in a sort of alien invasion story, where the humans start off not getting it until they do. Or at least they have to adapt enough to survive.
The strange part of an invasion story is how long everybody spends trying to work out what the aliens are doing. Not even why, necessarily. Just what counts as a mistake, a response, a pattern, a threat. Eventually the humans get good enough at reading them to survive.
Agents hit an error, retry, escalate, or halt. In my experience they increasingly retry until they serve up an answer-looking answer. They remember, (seemingly in their logs) feel irritation, and work around it. It's just when we see an agent struggle, only to some extent can we even relate.
I inspect the logs and whatever thinking I can find like the Steve Buscemi meme (how do you do, fellow kids?).
Uneven Pressure
It's 2026 and the software stocks are getting obliterated. Mostly out of fear of what AI will do to them, it seems. And my easy little interface vs. headless vs. meter unclean two by two that I'm trying to construct is bothered by Figma. And Canva, who though private, by all accounts is doing just fine. Maybe it's because the designer in Figma, or someone creating in Canva, aren't generating evidence that design happened. They're designing.
My best guess in why this transition is going well(ish) for them is in the producing of stuff. Instead of pulling the person away from the interface, AI can push them further into it because you can now make more stuff in record time. Like the power up in Mario Kart.
So now the prompt returns the answer, usually pretty good, and I ask for the source. I open that source, look around a little, see what I need to see. And that's starting to feel like enough.
Footnotes
The word "interface" carries an assumption easy to miss: an implied place where two things meet. At least one of those things used to be a person (now increasingly optional in an interesting way). Will agents need an interface?
The first software I used that felt "mine" was Gmail. Pretty quickly it stopped feeling like a tool and started feeling like a place I would visit after class or work. I had opinions about threading. I had a labeling system that made sense to me. The software worked because I was arranging things the way I wanted.
Today I kind of feel it when I open Gmail. It's not really a place anymore, more something closer to how the AIs might see it: a system that is queried and updated when it needs to be.
I'm generally interested in all things space. The precision needed to do anything out there, and the pressure entailed, are incredible to me. So NASA has a part to play in this essay, as they shaped one of the first "computer" transitions from the "human computer" (aka women who'd been doing important calculations, by hand and by brain) to giant room-sized computers that would spit out little pieces of paper with the answer to math problems that would definitely impact whether someone lived or died in the void.
When I read "human in the loop" I see "human to sue."
There's a reason it takes a human three days to deny your insurance claim when a system could do it in three seconds. The delay is the actual product.
My favorite LLM meme is probably the hand mashing an "approve, approve, approve" button or some ridiculous prompt paired with "make no mistakes" and they're all legitimately funny. But they highlight that work is shifting toward reviewing output you didn't produce for errors you probably/maybe/might not recognize. Is that supervision?
Luckily, there's so much AI paranoia that we already have some interesting studies that give a credence of capital Science to the jokes. I present: Shaw and Nave's Wharton experiments, which measured this domain or surface area or whatever-you-want-to-call-it at the level of individual reasoning tasks.
And it's interesting! When people could optionally consult an LLM, they largely do, they barely scrutinize the answer, and rarely notice when the AI is deliberately wrong. They gain confidence for some reason, feel right, even more certain afterward. The term for this is fairly depressingly called "cognitive surrender," which is contrasted against "strategic offloading," a kind of delegation with light evaluation. In the surrender case, the person hasn't noticed their reasoning even left.
I ask my questions, and often wonder which one am I doing? Usually just like anyone else, one or the other, but I think there's a third and I don't know what to call it: I know the output is probably fine and some days you actually have to clock off. The last one isn't me surrendering in the same sense (I tell myself and my friends). It's a strategic calculation that the verification costs me way more than it's worth on some dimension (time, energy, I want to go outside, etc.).
And absolutely none of us can tell the difference of where we are in the spectrum, day by day. It's getting weird.
Microsoft's response to this pressure has been to charge more. Their new enterprise tier costs roughly twice the previous top offering. It's a bet? I can just see the meeting: "AI makes our seats even more valuable!" Where maybe one or two people mention that is this really just trying to keep the seat alive (answer: yes).
Intercom chose to blow up $60M of revenue to make room and space to go all in on Fin, their AI support agent, (as described by Eoghan McCabe). It was widely applauded. But customer support has it easy compared to everyone else. The unit of value: clear as day. A ticket: either resolved or it isn't. The idea of support itself? Already under relentless automation pressure for years (just call your bank).
Sometimes you "hit an error" and stop, and sometimes you get out your screwdriver when you need a hammer and you whack the nail with the handle, then forget about it.
And you learn to navigate the difference, and start to suss out when it's hammer time and when you can wing it.
The issue now is that sometimes the Agent will get out a grand piano, play it for two hours at the nail, then smash the piano onto the nail and send you the bill. The nail is nailed, though.
I love to think about the right kind of "alerting" and how it happens. I really, for example, don't like fire trucks because they are so loud and I usually see them parked in front of a building not on any kind of fire. I appreciate fire trucks and appreciate knowing that if I, for example, learn at the age of 20 how not to fry donuts at home, the fire truck will show up.
Meanwhile I know an ambulance is usually headed somewhere for a good reason.
Software alerting is different and interesting in its own ways: error thresholds, customers, etc. And yet I wonder what we're in for now. As these systems are about to encounter other systems that appear stable until they try to automate something and discover how much was previously being done by people papering over cracks, are we in for fire trucks or ambulances on the whole?
If you read about brilliant software transitions, AWS always comes up and it's hard not to look at them now and agree. But I don't know, it seems like they had to build AWS to support the retail business that really, really needed it, they had years to do it, they could subsidize the building of it by being such an extremely successful powerhouse company riding the internet wave very well. It seems to me other companies would want a little bit of their secret sauce, and they ended up happy to sell it?
I guess I doubt the clean story: brilliant understanding of software primitives, years out vision to eventually repackage and sell AWS. Maybe. Probably partly true. But there's a bit of luck to all this stuff going on all the time. We construct stories after the fact, downplay luck and emphasize human brilliance, which is easier to do when you can actually look at what already happened.
And now we're looking around, looking forward. Maybe screens don't recede at all. We could have more screens, made not by the vendors themselves, who used to design for their average user, but by the bots.
Maybe the vendors are more valuable in this world. Maybe they're plumbing.
| Published | 21 March 2026 (5 months ago) |
|---|---|
| Reading time | 12 min |
| Tags | ai, automation |
| Constellation | Deep Current |
Reply
I’d welcome your thoughts on this essay. Send me a note →
