My Phone Was the Org Chart

For a while this summer, the most overworked node in the Thrivbe network was my phone. Every automated process we run — and we run many — reported directly to me on Telegram the moment it had anything to say. A job finished: ping. A report landed: ping. Something actually broke: ping — same sound, same weight, same claim on my attention as the harmless ones before it.

I've spent a year writing about hierarchy as an information-routing protocol — the Roman decanus, the Prussian General Staff, the whole two-thousand-year workaround for the fact that one human can only track so much. Then I looked at my phone one morning and realised I had rebuilt the exact problem I keep writing about. No middle layer, so every signal from every process travelled straight to the top of the pyramid. The top of the pyramid was me, in bed, at breakfast, mid-conversation.

The same problem, smaller

The org chart problem and the notification problem are the same problem. Both come down to routing: who needs to know this, who needs to decide it, and how urgently. That triage is what middle management was actually for, before bureaucracy calcified it into status meetings and forwarding chains. Block is trying to rebuild that routing with AI for a company of thousands. I needed the same thing for a company of one person and a growing crew of agents — because it turns out you can drown in middle-management traffic without employing a single middle manager.

So in July, in a single day, we shipped three connected systems on our own AI operating system — it runs on a server we call Thrivbe-1 — to take that routing job away from my nervous system and give it to the infrastructure.

Two reasons to interrupt a human

The first fix was the simplest: a split. Routine, informational output — the job finished, the report is ready, all is well — now flows into a silent feed inside Command Center, our internal dashboard. Only two categories are allowed to reach my phone: actual failures, and things that genuinely need a decision from me. That's it. Two doors.

The system didn't get less informed, and neither did I. It got quieter. I read the feed the way you'd read a ship's log — on my own schedule, when I choose to — instead of the feed reading me. All the information is still there. What's gone is the assumption that every piece of it deserved an interruption.

The infrastructure learned to heal

The second system is the one I'm most proud of, and not for the reason you'd expect. When something breaks now — a service crashes, a scheduled job fails — the failure automatically files an incident, and the incident wakes a repair agent. The agent reads the logs, the failing code, the recent git history. It diagnoses. And if the problem sits inside a deliberately narrow fence — it may only edit the failing project's own code and restart that project's own service, never delete anything, never touch another service, never go near credentials — it writes a fix, tests it, commits it, deploys it, and verifies the fix actually held. If it isn't confident, or the fix doesn't hold, it stops and flags me instead of guessing.

We tested it live: deliberately broke something and watched the whole loop run. The moment that convinced me wasn't a successful fix. It was watching the agent hit a failure that fell outside its fence, recognise that, and escalate to me rather than pretend. The refusal was the proof. Anyone can build an agent that tries things. The one worth trusting is the one that knows when to stop.

Silence means consent

The third system handles the small decisions that used to interrupt me mid-flow — a config choice, a "should we do A or B, need your call." Those now get filed into a queue instead, each one carrying a sensible default and a deadline. I can answer any of them, any time, with a simple command. If a deadline passes without a word from me, the system proceeds with the default — silence means consent — and logs that it did, so nothing is ever silently stuck waiting on my attention.

The interesting effect isn't on me. It's on whoever files the decision, human or agent: you only attach a deadline if the default is genuinely safe to act on without me. A decision that can't safely default just waits, undated, as long as it has to. Filing something into the queue now forces the honest question underneath every interruption: what actually happens if Robin never answers this? Often the answer is "nothing bad" — which means the interruption should never have existed.

An immune system, not an inbox

Your body runs the same architecture. It doesn't page the brain about every pathogen. The immune system detects the failure, handles what it can locally, within strict bounds, and escalates only what it genuinely can't — the fever, the pain, the signal that this one needs the whole organism's attention. Detect, heal within scope, escalate honestly. That isn't a metaphor I'm decorating the system with after the fact; it's a functional description of what we shipped.

It's also what a healthy hierarchy was always supposed to be: route each problem to the smallest level that can absorb it, and pass upward only what that level truly can't. Somewhere along the way we calcified that into bureaucracy, and then into notification settings, and both ended up doing the opposite — escalating everything, resolving nothing.

My phone is quiet now. The system is busier than it has ever been. Those two facts are connected. So, honestly: of everything that will interrupt you today, how much genuinely needs you — and who, or what, is doing that triage?