I Built an Executive Team of 6 AI Agents to Manage My 15 Side Projects

I have 15 active side projects and a full-time engineering job.

The math doesn't work — unless you delegate.

So I built an executive team of 6 AI agents. Each one has their own domain, personality, and skill set. I call it my Executive Cabinet.

The Team

My Role as Commander

I focus on four things only:

• Strategic decisions — what to build, what to kill
• Being the face — presentations, networking, relationships
• Building relationships — partnerships, collaborations
• Validating ideas — testing hypotheses with real users

Everything else is delegated. Coding, inbox processing, blog posts, goal tracking, competitor research — all agents.

How It Works in Practice

I send a message (usually via Telegram). Maya triages it by domain and routes to the right agent.

A typical day:

• Morning: Maya + Sage run daily review and set priorities
• Midday: Viktor reviews PRs, Luna drafts content
• Evening: Maya generates a report, Marco checks weekly goals

Each agent runs in its own tmux session with an isolated git worktree. They deliver pull requests, not just local commits. Everything is reviewable.

The Tech Stack

Nothing proprietary. No custom platform.

• Claude Code (Opus) — the brain
• Markdown files — skill definitions, prompts, context
• Git worktrees — isolation per agent task
• Tmux — parallel agent sessions
• Notion — Kanban board for tracking
• Telegram — input interface

Governance: Sociocracy 3.0

The team follows S3 patterns:

• Clear domains — each agent owns a specific area
• Consent-based decisions — no one overrides another's domain
• Driver-based work — every task starts with "why" (tension, driver, requirement, response)
• Accountability — agents must deliver PRs, not just status updates

What I Learned

1. Agents need structure, not freedom. Vague prompts produce vague results. Each agent has a detailed skill file with step-by-step processes.
2. Fire-and-forget beats micromanagement. I dispatch tasks and check results later via a /scrum command that reads all agent logs.
3. The inbox pattern is everything. One command (/inbox: <task>) creates a worktree, writes a prompt, launches in tmux, and logs everything for retry.
4. Personality matters. Giving agents names and domains isn't just fun — it creates clear routing and accountability.

Try It Yourself

The whole system runs on Claude Code with markdown skill files. No special infrastructure needed. Start with one agent (a Chief of Staff for daily reviews) and expand from there.

I recorded a 7-minute video walking through the full setup, presented from my Apple Vision Pro workspace. Watch it above or on YouTube.

What's your approach to managing multiple projects with AI? I'd love to hear how others are doing this. Find me on X or drop a comment on the YouTube video.