How a Recursive Agentic AI Organization Actually Works

01The Org Chart

I structured my one-person business into a 13-department AI organization. Each department has a clear scope, its own specialist sub-agents, and its own memory layer. The whole thing runs on Claude Code with opinionated CLAUDE.md files that route work to the right place.

Each department has its own CLAUDE.md defining scope, tools, voice rules, and constraints. When work routes there, only that department's instructions load — the others stay quiet. No context bleed. No instruction overload.

02The CEO Dispatch Protocol

Every incoming idea, request, or interruption goes through the CEO. The CEO doesn't do the work — the CEO classifies it. Three routes:

This is how solopreneurs stop drowning. Every email, every Slack message, every "hey could you also do X," every shower-thought idea goes through one classification function. The system tells you what to do with each one. No more "I'll get to it" lists that grow forever.

03Sub-Agent Spawning

When a department picks up work, it spawns a sub-agent with its own isolated context, role-specific permissions, and a clear scope. The marketing copy you're writing doesn't leak into the engineering review. The financial reconciliation doesn't burn context on the proposal you're drafting.

What this means in practice:

• Your main context stays clean. The dispatch hub tracks status. Sub-agents do the work in their own context windows.
• Each sub-agent gets only what it needs. A code review subagent doesn't see your tax records. A marketing copy subagent doesn't see your CAD files.
• Parallel work becomes easy. Three sub-agents can run simultaneously — competitor research, blog drafting, schema audit — without stepping on each other.
• 100X output, not 100X cognitive load. You orchestrate. The org chart executes.

04The Memory Hierarchy

Memory at the right layer means nothing is forgotten — but nothing overflows either. Four layers, each loaded only when needed:

The result: your AI knows who you are at every level — globally, domain-specifically, project-specifically — without ever loading more than it needs to. Memory that compounds without bloating.

05The Skill Ecosystem

50+ proven techniques across departments. Each skill is a documented capability the system can invoke — no reinventing the wheel, no rebuilding from scratch every time.

The flagship: gstack — a virtual engineering team covering CEO review, eng manager planning, designer consultation, QA testing, security audit, and release engineering. Every build runs through it. The standard build loop: /office-hours → /autoplan → build → /review → /qa → /ship.

Plus domain-specific skills for marketing (copywriter), design (visual storyteller stack), finance (trading systems), engineering (CAD readers, BOM parsers), and more. The system doesn't reinvent the wheel — it spins one you've already validated.

06Hooks, Slash Commands & the CLAUDE.md Cascade

The operator vocabulary that makes the system speak fluently:

Hooks are automated triggers — code that fires before or after specific tool calls or session events. Want a memory write to happen automatically when a session ends? Hook. Want a backup to run before a destructive command? Hook. Behavior the system enforces without you having to remember.

Slash commands are operator vocabulary. Instead of explaining what you want every time, you invoke a known capability: /office-hours for CEO review, /qa for browser testing, /ship for release. Each command has a documented contract — same input, same output, every time.

The CLAUDE.md cascade layers instructions: global (your identity) → project (the codebase) → department (the dept scope) → sub-department (specialist context). Each layer adds opinions without overriding lower-priority instructions. Together: an operator vocabulary the system speaks fluently.

07What Anthropic Ships vs What PrimoLabs Ships

When Anthropic shipped persistent memory for Claude Managed Agents in April 2026, three people asked me if it commoditized what I'm building. Here's the honest comparison:

Anthropic ships substrate. PrimoLabs ships the opinionated organization on top. Both grow together — one doesn't replace the other. Read the full argument →