Personal Hub
A personal agentic workspace, and how it runs.
Running agents turned out to be the easy part. Keeping track of them was the hard one.
By October 2026 I had work going on in Claude Code, in Codex and in the Claude app, and notes sitting in memory folders all over my machine. None of them knew about the others. Every new chat started from zero, so I'd copy context out of the last chat and paste it into the next one.
The structure came out of a voice note. In July I recorded myself thinking out loud about agent workflows and which parts of them are here to stay. It ran long. I started with what I was sure of.1The recording is from 3 July 2026. I pasted the transcript into a chat. Typos are fixed here, and the dots mark where I cut.
Agents are not going anywhere. They're here to stay. Skills are not going anywhere. They can be redefined into something else, we don't know yet. Loops, they're here to stay. Workflows, they're here to stay. … It starts with an army of agents. And an army of agents is basically what we call a workflow.
Then it turned into an org chart. Every company has a CEO, and everyone at the bottom reports to someone one level up. In my version every person runs their own agents, the way a country runs itself.
A country works for its own best interest, so a person's agents should work for theirs. That's a reward function. For most people the terms are the same: money, work they love, fun, family, health. What changes from person to person is the weight, the beta, on each one.
Maybe the beta coefficient for having fun is higher for some people, and it's lower for some people. And the beta coefficient for working and earning that reward is higher for some people, and it's lower for some people. Some people just want to inherit, and some people want to just get some money from a lottery or something like that. And just enjoy life.
R = β1 money + β2 work I love + β3 fun + β4 family + β5 health
In the hub the reward function is a real file, and any agent that has to rank my work reads it first. The agents are how I push on all five terms at once.
These armies of agents are nothing but workflows. Every workflow is reporting to another workflow, and that workflow is reporting to another workflow, and at the top, I am there.
Nine minutes later I typed the short version into the same chat.
I'm the CEO. I can't drill down into every little detail, I can only manage it from the overview.
Talking it through out loud is what gave me the framework. Sessions became the departments, agents became the staff, and every result climbs back up to me. It took until October to build it.
i. The org chart
Everything lives in one private git repo, with a folder for each part of my life. Each folder has a short summary. Every session reads the top index before it does anything else, then reads only the summaries its job needs.
Sessions form a tree. The hub session sits at the top, then one session per area, then one per project, then the sessions that do a single job. Any session can start a child with /spawn.2It opens the child in a new terminal tab, as a copy of the parent's chat (a fork), or in the background (an auto session). One folder can run five auto sessions at a time. When the child is done it sends one to three lines back up, so the parent never has to read the child's whole chat.
A small app shows the whole tree live in its own tab. A session that's waiting on me gets marked, so I look where I'm needed and leave the rest alone.
ii. The gate
Agents can propose whatever they want. A script that gives the same answer every time decides what passes, and I decide what goes out.
If an agent writes code, a second agent reads it before anything runs on it. If it writes something in my name, a script checks it against a list of phrases I'd never use. And nothing gets sent, posted or committed until I say yes.
iii. Memory
Twice a day, at 03:00 and 13:00, a sync runs by itself. It reads every Claude Code and Codex transcript, plus the exports I pull from the Claude app, and keeps only the words. Each chat gets filed under its area. Then an agent updates the summaries from whatever is new and pushes the repo.
The next session starts from the summary. If it needs a detail the summary dropped, /recall3A full-text index of 1,392 chats on 12 October 2026, the oldest from February 2025. The sync rebuilds it on every run. searches every chat since early 2025, more than 117,000 messages, and answers with the chat it came from. A search takes under a second.
iv. From my phone
The PC stays awake on power, and four remote servers start when I log in, one for each main folder. The Claude app on my phone connects to them. From a session on the phone I can /spawn into any folder, and the child opens as a tab on the PC and shows up on the phone too.