2024 – present
Agent fleet and chief of staff
A chief-of-staff agent routes work to about 30 specialists across four businesses, and a person approves anything that sends, publishes, or spends.
Private repo, walkthrough on request
- Specialists
- About 30
- Businesses
- 4
AI details
What the system actually does
Models
- Grok, Claude, ChatGPT/GPT, Gemini, Perplexity, Kimi K3, and Microsoft Copilot Cowork, the models I've used hands-on.
- A lighter, faster model is the default for routine work.
Techniques
- Chief of staff routes each request to the specialist who owns that function.
- Scheduled routines and event triggers (email, webhooks) run recurring work.
- Human approval before send, publish, or spend, with the draft shown first.
- Credentials stay in an encrypted vault and are filled in without the model seeing them.
- Allow-rules cover a short list of actions, and a safety review blocks the rest.
Data
- Inbound email, calendar context, and Drive files through Google Workspace MCP connectors.
- Operational work for Follow Apparel, US Buy Direct, Pathwords, and Mathteroids.
Evaluation
- Multi-agent QA is required before anything ships.
- A weekly routine healthcheck audits scheduled automations for wasted tokens.
- Standing rule: stay silent when nothing changed.
- Routines that no longer earn their cost are paused or deleted.
Challenges
- One front door has to cover email, research, calendar, copy, design, QA, SEO, monetization, Amazon, and research without mixing their rules.
- The model must not see credentials or send, publish, or spend on its own.
- A large scheduled fleet gets expensive if empty runs still report.
Innovations
- About 30 specialists behind one chief of staff, with per-business owners and shared rules.
- Email triage drafts a reply and stops. Nothing goes out until a person approves the draft.
- Work waiting on a person often parks in Project & Task Command Center.
Stack and tags
- MCP
- Gmail
- Google Calendar
- Google Drive
- Webhooks
- Cursor
- Vercel
- agents
- approvals
- mcp
- operations
Problem
Operational work for four businesses was coming in as email, webhooks, schedules, and one-off requests. It needed one front door, a specialist for each kind of work, and a person still signing off on anything that leaves the building.
Why I built it
The fleet does real e-commerce, mobile app, and content work for Follow Apparel, US Buy Direct, Pathwords, and Mathteroids. It isn't a demo. Every business shares the same rules and has its own owners.
How it works
It's a coordinated team of about 30 specialized agents. The chief of staff routes work to the specialist who owns that function: email triage, intel and research briefs, calendar, copywriting, design, QA, SEO for AI answer engines, monetization, Amazon, or research. Recurring work runs on scheduled routines and event-driven automations, so nobody has to start it each time.
Email is one of those paths. Agents connect to Gmail, Google Calendar, and Google Drive through MCP connectors on Google Workspace. The chief of staff routes a message to the email specialist, which classifies and summarizes the thread and drafts a reply when one is needed. Calendar mail can lead to a prepared hold, but no meeting gets booked on its own. No reply goes out without explicit approval, and the draft is shown first. Anything that would spend money or publish goes through the same gate.
A person approves anything that sends, publishes, or spends money. Credentials stay in an encrypted vault and get filled into pages without the model seeing them. That work often waits in Project & Task Command Center.
Written allow-rules cover the few actions agents can take on their own, like certain Vercel deploys or Drive uploads. An automatic safety review blocks anything else risky. Nothing ships without multi-agent QA.
I treat token use as a cost. A weekly routine healthcheck audits the scheduled automations. The standing rule is to stay silent when nothing changed and to use a lighter, faster model by default. Routines that no longer earn their cost get paused or deleted. A review of repeated friction turns that work into reusable skills, so it isn't redone from scratch.
The models I've used hands-on are Grok, Claude, ChatGPT/GPT, Gemini, Perplexity, Kimi K3, and Microsoft Copilot Cowork, plus Cursor.
Outcome
The same pattern runs daily across all four businesses. Inbound mail gets triaged and drafted inside that workflow, and a person still owns anything that goes back out. The fleet can keep running without paying for empty reports.