Design the System Around Your Attention, Not the AI
Asif AhmedAgents can work all day, BUT I can’t watch them all day. I am not at computer screen OR want to be glued to my phone all day. However, I do want things to be working in the background as new Slack message or new emails come in. I’m limited by my ability to pay attention and make decisions - I don’t want to outsource the most critical thinking aspects. So I built a system around those limits.
1. Attention is the constraint
In a normal week I drive a few dozen AI sessions by hand. More than 200+ run in the background, building, checking, reviewing and reporting. I drive certain sessions more by hand and certain sessions do not need me to babysit them. But when I come back to my computer, I need a list of decisions based on the domain. At any given point there are 10+ sessions running actively OR in the background. It gets hard to know what’s going on in each at a granular level - attention becomes the constraint!
2. Give every kind of work a lane
Let’s say you only had one long chat with one AI. It would break down fast because each session can only really work on one thing at a time. Yes, subagents exist but it’s not quite the same thing. AI forgets certain aspects AND it mixes topics. It’s going to be better to try to come up with background sessions that have their own domain.
Each domain has its own long-lived session with its own queue, like specialists on a team: the product, fund data, contacts, the mac app, the phone, specific new features. A manager session sits above them. It takes a request, sends it to the right lane, and checks the proof when the work comes back. It can’t ship anything itself.
Lanes keep each agent’s context small and its job clear. They also mean a lot of agents want something from me at once.
3. One front door for decisions
If every agent can ping you, you’ve built a very fast way to get interrupted and that is not very efficient. You only want one interface to these background agents.
So every question for me goes to one inbox, and each one comes with the agent’s recommended answer and a real example: “I’d do X, here’s the actual item, here’s what could go wrong.” Often when I’m away from the computer, I say “I’ll be back at 830AM, make sure to have my decision doc lined up”. Then I usually rapid fire on decisions, but about 25% of them, I’ll need to dig deeper or get more context. This type of calibration is going to be a work in progress. You don’t sharpen your tools only once, and similarly you need to sharpen your AI system always.
4. Let work come in through any door
My best ideas rarely show up while I’m at a keyboard, so the system takes work from wherever I am. I type a short nudge into a session, or cut in while one is still working. Almost a fifth of my messages land mid-task, and steering costs less than restarting. When an idea needs a paragraph, I dictate it; voice is a small share of my messages but a big share of the words. An emoji on a Slack message becomes a to-do that clears itself once I reply (I’ve been tagging a lot of messages with an hourglass emoji recently). From the car I dictate via voice, and it files the ask for me. On my phone I tap answers in an inbox UI.
Each of those ends up as the same work item: what to do, why, and what done looks like. The system treats them all alike, and I get to use whichever door fits the moment.
5. Draw the trust line on purpose
More doors and more agents mean more ways for the wrong thing to happen, so I drew a line.
Inside it are the agents that run on my own machines and my own account: my personal assistant on the Mac and phone, the coding sessions, and a planning agent. They see my inbox and act on my word. Outside it are the voice agents I talk to on the go. They’re fast and useful, but they read only messages addressed to them and never touch work data.
Three things always come back to me, whoever asks: shipping to production, spending money, and sending anything as me.
Because the boundary was there from the start, adding another agent becomes an easy call. Each new one starts outside the circle of trust.
6. Ask for proof
Proof and verification are super important. Agents say “done” all the time (aren’t they little fibbers sometimes?). Therefore work goes through three checks before it reaches me. And I NEED a visual usually to validate what “done” looks like.
The first is tests, few and sharp. Hundreds of tests that only check the code agrees with itself tell me little. I want one test for each real way the thing has broken or could break, written so it fails without the fix. Changes people can see get checked the way a person would use them: a real browser clicks through the page and confirms the result on screen. After a change goes out, a check reads the live state back. A green test suite is where checking starts.
The second is review by someone who didn’t write it. The agent that builds a change never grades its own work. Separate review agents read the diff, each hunting one kind of problem: a correctness bug, an error that fails silently, a simpler way to write it. Then a reviewer session compares every claim in the hand-back (“tests pass”, “it’s live”) against the evidence attached and sends back any claim that has none. For a surprising finding, a second model from a different lab weighs in before anything changes.
The third is pictures of what’s live. Every change someone can see ships with before and after screenshots of the real product, plus a short video of the new behaviour. The deploy takes them itself: the change names the page it touches, and the pipeline shoots that page before it ships, again once it’s live, and records the clip. All of it lands on one page with the change, the tests and the review.
So when I open my morning page, I look at screenshots. I can trust a result in seconds, and when something looks off, the evidence shows me where.
7. It compounds
Every correction becomes a rule, a test or a check, written next to the work it guards, and the next agent inherits it. I don’t have to teach the system the same thing twice.
It started small, as a phone voice assistant in late summer. By early autumn it was a personal agent on the Mac and the phone. By mid-October it had lanes, and soon after that a manager. Each step took one more thing out of my head.
Start with one lane and one front door
You don’t need a fleet to start. Pick the kind of work you repeat most. Give it its own session and a clear definition of done. Send all its questions to one place, each with a recommended answer, and ask for evidence with every result.
Then watch where your attention still goes, and build the next piece around that.
