How a 17-agent AI chief of staff is organized.
My AI chief of staff is one supervisor agent, Rory, that takes in everything asking for my attention and routes each piece to an agent that owns that one job. The agents share a change feed, every scheduled run writes one line to a log, everything lands on one board, and permission rules stop anything leaving in my name until I say yes.
I started with one assistant and added each piece below when I needed it. Here's the structure, drawn out, with what each piece is for. It runs on Claude, and none of it needed a server of its own.
Daily operations
- Morning brief and evening close
- Inbox triage
- The running list of open items
- Writing in my voice
- Weekly review
The ventures
- Builder
- Product and roadmap
- Marketing and growth
- Design
- Customers and fulfillment
- Portfolio priorities and review
Home and life
- The house
- Kids
- Menu planning
- Cleaning routines
- Kid schedules
What does the supervisor agent in an AI chief of staff do?
The supervisor is the one agent you talk to. It reads what comes in, decides which agent owns it, hands the work off, and keeps the rules every agent follows. It passes the other agents' results back to you without redoing their work.
Mine is Rory. When I ask for something, Rory hands it to the agent that owns that job and brings the result back to me. Because the supervisor doesn't do the work itself, I can add an agent without rewriting the others.
How do the agents in an AI chief of staff coordinate?
Through two shared files: a change feed that says what changed, and a registry that says what's true. Every agent reads the feed before acting and posts one line after any change another agent would be wrong not to know about.
The registry holds the IDs and settings everything depends on, and when an ID pasted into an instruction disagrees with the registry, the registry wins and the mismatch gets flagged. In the 30 days before my September audit, the agents posted 332 lines to the feed.
What runs on a schedule, and what runs live?
Briefs, reviews and reports run on a schedule; anything that needs a decision runs live with me. Scheduled runs follow stricter rules: they never ask a question, never take an outbound action, and fail loudly, naming what broke.
A morning run builds my brief before I'm up and an evening run closes the day. Weekly runs pull the numbers for my sites on Monday, write a marketing report, and review the week on Friday. Live sessions are where I approve things, build things and change my mind.
How do you know an AI chief of staff is actually working?
Every scheduled run writes exactly one line to a run ledger, a separate check alerts you if the morning run hasn't finished by a set time, and a weekly self-grade compares the ledger against the full list of runs that should have happened and names the absentees.
My first version stopped running in May and I didn't catch it for three weeks. Now every scheduled run reports when it fails, and I get an alert if the morning run hasn't finished.
What tools does a multi-agent AI chief of staff need?
One AI platform that can run agents with instructions and on a schedule, a durable place to keep records, read access to your email and calendar, and somewhere to see it all. Mine is Claude Code and claude.ai, Notion as the record, my email, calendars and messages as read-only inputs, and a simple board that reads a file the agents write.
The board only shows what's in the record and doesn't store anything itself, so if the two ever disagree, the record is right. Each agent's instructions live in one folder under version control, so I can see and undo every change to how an agent behaves.
If you're curious what your version would look like, book a free 30-minute call or read how to build one, step by step.