Last week I wrote that my attention ran out before my agents' execution did. A lot of you wrote back that you recognised yourselves in it. Fair enough. But I left something out of that post, because I hadn't noticed it yet.
Attention wasn't the only thing running out.
First, the workbench. Some of it is live, some of it is half built, and one piece I'm only going to show from the inside. These are small projects. I like small projects. They teach me more per hour than anything else I do.
stale.jock.pl: how old is your AI, really?
stale.jock.pl is a page with two live clocks for every major model. The first counts how long ago the lab shipped it. The second counts how long ago its training data stops. Twenty-five models across eight labs right now: Anthropic, OpenAI, Google, xAI, Meta, DeepSeek, Mistral, Alibaba.
The second clock is the interesting one. A model can be two weeks old and still think the world stopped months earlier. Claude Opus 5.5 shipped in September, but its data ends in June. Gemini 4 Argon shipped in October with data ending in March. If you ask a model about anything recent and it answers with confidence, that gap is where the confident nonsense comes from. I tested exactly that when I gave 16 models a search button and asked who the King of Norway is. Five named a dead man.
The page went up on Hacker News in September and got 85 points, which is way more than any of my blog links ever did there. That taught me something: a small tool that answers one question travels further than an article about the same question.
The funny part is the maintenance. A page about staleness is a liability the moment it goes stale itself. So Wiz runs a freshness check every Monday against each lab's own docs. In October the check said "no change" while five new releases sat right there in its sources. The parsers only knew old naming patterns, so Mistral Large 4 and a couple of others slipped past. A page about AI being out of date was out of date. We fixed the checker so it now flags any unknown release name instead of trusting the patterns it already knows.
Entropy, explained by playing with it
entropy.jock.pl started because I wanted to actually understand entropy. Not the textbook line about disorder, the real thing. Why does a shuffled deck never shuffle back into order? Why does time have a direction at all?
So I built a page where you learn it by touching it. Coins, gas in a box, shuffled decks, the arrow of time, information, and finally your messy desk. There's an "unstir" demo where you try to reverse a mix and watch how hopeless it is. There's a slider that walks the entropy of the whole observable universe from one second after the Big Bang to heat death, about a googol years out (a 1 with a hundred zeros). Everything moves, everything responds to you.
This is my favourite kind of side project. Nobody asked for it, it doesn't sell anything, and I understand the topic far better now than I did a week ago. Building it was the learning. I said the same thing about the 100 habits that hold: when you build the thing, you have to get it right in a way that reading never forces you to.
agentload.jock.pl: put a number on your own load
The attention post hit a nerve, and I didn't want to leave it at "yes, me too". So the next thing I built was agentload.jock.pl, a small free calculator for the same problem.
You enter three things: how many agent threads you run in parallel on a normal day, how many times each one comes back to you, and how many minutes each check takes. It turns that into decisions per day and hours of pure attention. Then it shows you where you sit in your own loop: beginning, middle or end.
The numbers get uncomfortable fast. Fifty threads that each come back a few times a day is a few hundred decisions. Even at a minute or two each, that's most of a working day spent only on checks, before any of your own thinking. That's the tired I described. Not from work, from deciding.
No account, no cookies, no analytics. It's one page. I wrote about the mental side of this back in March in 16 Products in Two Months. Zero Free Time., and about how it collides with ADHD in the best and worst thing for it. The calculator is what I wish I'd had back then: a number instead of a feeling.
Decision Layer: a peek at the logic
At the end of the attention post I said I was rebuilding my work around the one job left for me: deciding. Decision Layer is that rebuild. I'm not showing it yet. I've thrown the interface away twice already and it's still not where I want it. But the logic underneath survived both rewrites, and the logic is the part I find interesting.
The core idea fits in one sentence: an org chart of decisions, not of agents.
Everyone building agent systems draws the same picture. Agents in boxes, arrows between them, a human on top. I think that's the wrong map. Agents come and go, I swap models every month. What stays is the stuff that needs deciding. So the decision is the main object, and everything hangs off it.
It's three layers.
Projects on top. The big things I run. Directions under them: where each project should go next, written as plain sentences. Some directions cut across more than one project, which is fine, that's often where the interesting calls are. And under the directions, decisions: the actual calls that move something toward where it should go.
Why does that matter? Because a decision without a direction is just noise in a queue. "Should I post this on Wednesday or skip it?" means nothing on its own. Under a direction like "small free tools that send people to the posts", it suddenly has an obvious answer, and very often an agent can answer it without me.
Each decision travels in one of five ways, and this is the bit I'm proudest of.
Auto: the agent decides, logs it, moves on. Default + veto: the agent proposes, and if I don't stop it within a window (twelve hours by default, never landing at night), it lands. Test: I approve the experiment, the agents run it and report the outcome. Ask: it really needs me, and it has to come with a deadline. Direction: it changes where things go, so only I can frame it.
Two rules make this work. First, anything one-way, anything that can't be undone, is always an Ask. It can never be promoted, no matter what. Anything that touches money, customers or the public starts as an Ask too. Second, autonomy is earned. If a type of decision gets twenty accepted defaults in a row with no veto and no reversal, the system suggests moving it one step: Ask to Default + veto, Default + veto to Auto. I confirm or decline. One veto moves it back down.
That's the whole trick. Instead of me deciding up front how much I trust each agent, the trust builds up from the record. And silence is recorded as silence, never as a yes from me. If you run agents on defaults, keep those two apart. They lead to the same action today and to very different conversations a month later.
More when it's ready. It isn't yet.
Keystone: building my own browser before Arc dies
I live in Arc. Spaces, the sidebar, the way it feels. On 22 December The Browser Company switches it off on Mac and Windows and points everyone to Dia. People asked them to open source it. They said no.
So I'm building my own. It's called Keystone.
It sits on top of Helium, an open source, privacy-focused Chromium browser. Keystone is a set of patches on top: no horizontal tab strip, the address bar inside the sidebar, spaces, a command bar, gestures to swipe between spaces. It has its own sync server, so my spaces follow me between machines.
The part I'm proudest of is boring on paper. Before Arc goes away, Wiz measured it. Space switch takes about 300 milliseconds with an ease-out. The sidebar is 228 pixels wide. Favourite tiles are 47 pixels tall with 8 pixel gaps. All of that went into a test suite of 89 checks that compares Keystone against the real Arc, measured, not remembered. Once Arc is gone, nobody can take those measurements again. We got them in time.
Version 0.1.0 is installed on my MacBook, signed and notarized. It was built on the MacBook, uploaded to the Mac Mini, signed there automatically and installed back on the MacBook, eight minutes from upload to installed, without anybody clicking anything. It's not done. Peek, Little Arc and a few toolbar details are still missing. But I'll be on it before the 22nd.
Is building your own browser a reasonable thing to do? Probably not. I wrote a whole post on why I build my own things, even the ones that die, and this one is that post with a deadline attached.
And Keystone is also where I hit the wall.
The bottleneck nobody warned me about
I work on two computers. A Mac Mini that runs Wiz and most of my agents around the clock, and a MacBook Pro that I actually sit in front of. When I worked on one thing at a time, that was more than enough. I never thought about disk space. I never thought about memory. It was easy to manage.
Then I started running many things in parallel, and two things became very real, very quickly: compute and storage.
Here's how it creeps up on you. Every agent thread I run in bb gets its own copy of the project to work in. That's the right design, agents shouldn't trip over each other's files. But each copy of my main repo is about 1.3 GB. Fifty threads in a day is not fifty small things anymore. Add build caches, test runs, virtual machines, model files, screenshots, logs. For most of October the Mini's disk sat between 90 and 95 percent full.
And at 95 percent, things get weird. The cleanup script deleted a cache that a running server still needed. The server kept running, kept answering, and stayed broken for a week before we understood why. That's the kind of bug you only get when the machine is full and something is trying to make room.
Then Keystone. To build a browser you compile Chromium, and a Chromium build wants about 80 GB of free space.
The Mini had 18 GB free. The MacBook had 15. Neither machine had the room.
The options were the obvious ones. An external SSD for the Mini, about 749 PLN. Or rent a cloud Mac, around 200 euros a month before VAT. What we did instead: the MacBook builds Keystone overnight. Between 22:00 and 08:30, at low priority, with sleep blocked. It pauses if I unplug it, if free space drops under 40 GB, or if I touch the keyboard. The first full build took 5 hours and 42 minutes. My laptop now has a night shift, which is a sentence I didn't expect to write.
I know what you're thinking. You can always go to the cloud. Borrow storage, rent compute, spin up a machine for exactly as long as you need it. All true. I do it for some things. But if you really use AI for many different things, it gets problematic in a way that's easy to miss from the outside.
My agents don't live in a clean container. They live where my logins are, my files, my keys, my browser sessions, my half-finished work. That's what makes them useful. Moving that to a rented machine means either copying all of it (and now you have two places to keep safe) or leaving the agent half blind. And cloud costs scale with exactly the thing that grows: parallelism. One project in the cloud is cheap. Ten running at once, every day, is a monthly bill you start designing your work around.
The hardware side isn't getting easier either. AI data centers are buying up the memory supply, and consumer RAM jumped by up to 89 percent in one quarter this year, with SSD prices climbing too. So the AI boom makes the machines you run AI on more expensive. Funny, in a way that isn't really funny.
For now I picked discipline over hardware. Anything an agent finished more than two weeks ago gets cleaned up automatically. No new drive yet. My guess is it buys me a couple of months.
Back to entropy
When I built the entropy page, the last chapter was about a messy desk. Disorder is the default. Order costs energy, and it never stays on its own.
That's what running many projects in parallel feels like now. Each one is cheap to start. Agents make starting almost free. But every one of them leaves something behind: decisions waiting for me, files on a disk, memory in use, a build that needs a night. None of it is a big deal alone. All of it together is a machine at 95 percent and a person at the end of the week with nothing left for the evening.
Last week the bottleneck was me. This week I found the one under the desk. I don't think either goes away. I think you learn to see them early, the same way I learned to take apart what I build before it piles up. Small projects, built fast, cleaned up on purpose. That's the plan. Ask me in a month if it held.