Last Tuesday at 10pm I had twenty terminal tabs open and three of them were running Claude Code. One was waiting on a yes-or-no answer that had scrolled off-screen an hour earlier. One had finished and I couldn't tell which. One was still working. To find out which was which, I had to click through all twenty.
Three agents. Three. And I'd already lost the thread.
That's the embarrassing part. The people I learn from in this field don't run three agents. They run fifteen. If three can fog over by 10pm, fifteen isn't a workflow - it's weather.
Now walk into a coffee shop. Within three seconds you've parsed the room: who's working, who's on a date, who's waiting, who's stuck talking to someone they're trying to escape. You didn't read a list. You glanced. That's the move your brain is built for, and it's the exact move every tool we've built for managing agents takes away from you.
The agents got smarter. The interface didn't.
Now picture this instead
Three of your agents, in an office you can see. The one refactoring the auth module is at the whiteboard - still planning. The one writing tests is at its desk, typing. The one updating the docs has walked to the server room to run a build.
An hour later you glance up from your own work. Auth has moved to the outbox: it shipped a PR. Docs is back at its desk with a small raised hand: it's asking you something. Tests has wandered to the break room: done, nothing left to do.
You glance up. You don't read anything. You see it.
The rest of this essay is about why that one move - glance, don't read - is the whole game, and why nothing we currently use lets you make it.
Places are free. Lists are expensive.
The reason you can parse a coffee shop in three seconds is that your brain wasn't really working. There's a part of you that handles spatial information - where things are, where they were, where you've been - and it runs for free. It's the reason you walk into a house you haven't visited in fifteen years and still know where the bathroom is.
The researchers who mapped how that works won the Nobel Prize in 2014. The short version: the part of your brain that holds facts has a famous ceiling - about seven things at once, beyond which it overflows. The part that holds places doesn't. It's bottomless, and it's cheap.
The Greeks knew this twenty-five hundred years ago. A poet named Simonides built the first memory palace - put the things you want to remember as objects in the rooms of an imagined building, then walk through to retrieve them. It works because it converts a working-memory problem (hard) into a spatial-memory problem (easy).
Now look at how we manage agents. Twenty terminal tabs. A vertical list of session names. A dashboard with status badges grouped by status. Every interface we've built pays the hard-memory tax, and none of them touch the cheap system. We're doing mental arithmetic when we could be walking through a building.
Every interface has been a metaphor
You can tell a metaphor is failing when it starts to crack.
The command line was a typewriter - honest about what computers did in 1970, broken the moment they did more than type. The desktop was a desk: Xerox PARC looked at how office workers actually behaved, noticed humans don't process information in streams but lay it out - papers, folders, a trash can - and built that into pixels. It's the last time a mass-market interface used the spatial system on purpose, and it ate the world. The web was a page; we traded space for unimaginable scale. The smartphone was a slab of glass with apps as tiles - spatial at the resolution of which app, then back to a page once you're inside one.
The interface we're using right now is conversation. ChatGPT made it the front door of AI, and for one person talking to one model about one thing it's close to perfect. That's why it spread faster than any consumer product in history.
But a conversation is just a list of turns, and a list holds one thing in mind at a time. You don't run ten conversations at once. Even at a dinner party for eight, you track who's sitting where with the cheap spatial system and talk to one person at a time. Conversation was never built for ambient awareness of many concurrent agents - and that's the exact load we're putting on it now.
What changed, fast
In February 2025 Andrej Karpathy posted a tweet about vibe coding - the new style where the model writes most of the software and you barely look. The term went mainstream in months. And within a year, writing code with AI went from a faction to the default: the big developer surveys now put adoption near-universal, with a large share of new code on GitHub already machine-written.
You know all that. Here's the part that gets less airtime: the number of agents per developer climbed faster than the quality of any single agent did.
Boris Cherny, the engineer who built Claude Code at Anthropic, talks openly about his own workflow, and the number he keeps repeating is five plus five to ten - five sessions running on his laptop, five to ten more in the cloud. Ten to fifteen instances of his own tool at once. The most experienced user of Claude Code on earth, asked how he uses it, answers: fifteen of them, in parallel.
The Claude Code team has been shipping for exactly this. /loop, which lets a session run unattended for days. Scheduled jobs that run on Anthropic's infrastructure with your laptop closed. And weeks before I sat down to write this, a single screen that lists every concurrent session grouped by needs input, working, done - what they called, accurately, the multi-session control plane.
That ship is the loudest signal anyone in this space will get this year. The team running the largest fleet of agents on earth looked at their own workflow and said: we need a better way to manage many of these at once. The problem is real, it's worse than most people realize, and it's the bottleneck.
But the way they solved it is still a list. A better list - status grouped vertically, the conversation metaphor extended one level up, where each conversation is now a row. The question nobody's asked yet is whether the metaphor itself is wrong.
The people who already solved this don't use lists
A friend who runs a small fintech told me he'd stopped giving his Claude Code sessions names and started giving them roles: "the migration one," "the auth one," "the docs one." The names were meaningless; the roles weren't. He didn't notice what he'd done. He'd started building an org chart in his head because his computer wouldn't build one for him.
Trading floors solved this thirty years ago. The Bloomberg terminal is famously hostile to look at, but every trader I've asked says the same thing: they don't read the screen, they remember where things sit on it. Currency pairs top-left. Equities in the middle. News scrolling on the right. Their fluency is spatial, not textual. The interface is ugly precisely because it had to optimize for one thing - tracking many entities at once with no time to read.
Air traffic controllers do the same. Surgeons do the same. Anyone who has to hold ten or twenty things in view and react in seconds eventually builds a spatial workspace, because the alternative doesn't scale. The list-of-conversations works for one or two agents. Past that, you need a room.
And no, a nicer dashboard doesn't fix it. Dashboards are good at forensics - you query them when you want to know exactly what happened at 3pm yesterday. They're terrible at ambient awareness. When you glance around a room you're not querying, you're noticing; information arrives without being asked for. That's a different cognitive mode, and a Grafana for agents - tiles, status lights, sparklines - doesn't deliver it.
Honestly, we should stop trying to bring people back to the office. The agents are doing most of the work now - they're the ones who need a desk.
What an office actually does
Go back to those three agents moving between rooms. Give it a name: status as location. Each agent's state has a place, not a string. The agent has a face you recognize, not a session ID. You take in everything in two seconds, without clicking anything.
This isn't just decoration. It's a skeuomorph that does real cognitive work - it takes a piece of hardware your brain already ships with (free, bottomless, evolved over millions of years) and uses it to carry the load your terminal tabs currently dump on your language system (expensive, capped at seven, exhausted by 8pm).
And it needs no new model. The models are extraordinary; the agents got smarter than most of us noticed a year ago. That was never the gap.
The bottleneck is the eight-inch gap between the agent and the human, and we keep trying to close it with a list.
That gap is what I'm building OfficeGlade to close: an office you can glance at, where your agents work, remember, and hand things off - and where you always know, at a look, which one needs you.
Why this is the moment
Every interface metaphor arrived when three things became true at once: the need existed, the technology got cheap, and the culture was ready.
The need is here. The most experienced people in the field are running fifteen agents each, and trust is sliding even as usage climbs - favorability toward AI coding tools has fallen for several years running in the largest annual survey. We don't need better models nearly as urgently as we need a better way to live with the ones we already have.
The technology is cheap. Browser-native real-time engines like Phaser and PixiJS let one developer build, in a weekend, the kind of spatial interface a games studio needed a year to ship a decade ago.
And the culture is ready. The last five years of mainstream gaming quietly retrained everyone - Stardew Valley, Animal Crossing, A Short Hike. Cozy spatial games are one of the fastest-growing segments in the industry, and pixel art has become a visual language an entire generation reads fluently. People are ready, in a way they weren't ten years ago, for a computer to feel a little like a place rather than a hammer.
One small thing before you go
In his 2026 retrospective on the original vibe-coding tweet, Karpathy said he was most excited about "the product of the two - better models and better agent layers." The model layer is in extraordinarily capable hands. The agent layer - where humans actually meet the models, day after day - is being figured out right now, by everyone reading something like this.
If you're running five Claude Code sessions across three projects tonight, you're not behind. You're early. You're in the room while the discipline is still forming.
The agents are already here. We're the ones still deciding where they live. If "an office, not a dashboard" is the right answer, I'd like you in the room while we build it - come see OfficeGlade, or get on the waitlist below.
- Natasa