Systems Architecture Case Study 13

Thirteen Agents.
One Operations Team of One.

A 13-agent Claude Code operations system, organized like a real marketing org, with a 5-layer shared knowledge library and a two-gate approval workflow that never auto-publishes.

13
Scoped, color-coded agents in production
5-layer
Shared knowledge library, read by every agent
2-gate
Approval (Bella in ClickUp + Aidan ship)
The Before

A one-person ops team doesn't scale by working more hours.

I'm a one-person marketing operations team at WHY Institute. Daily and weekly funnel audits used to mean pulling Mixpanel by hand, stitching it against ClickUp work-in-flight, cross-checking GHL email stats, opening four tabs in Slack, and writing the brief from scratch every Monday. Page production from request to live took days because every step (strategy, copy, design, code, deploy, QA) was a context switch. Email and SMS workflows lived in someone's head until I sat down and built them.

The work was getting done. It was just consuming every working hour I had, and the parts that needed the most thinking (strategy, copy diagnosis, business judgment) were the parts getting the least of my attention because the rest of the day was busywork.

Agent Org Chart · 13 agents, scoped like a real marketing team
CEO Aidan Final ship gate, never bypassed
Chief of Staff project-manager Agent 6 · blue
Strategy Director funnel-strategist Agent 2 · magenta
Reporting Director funnel-gary-reporter Agent 5 · yellow
Auditor funnel-auditor Agent 1 · cyan
Page Editor funnel-page-editor Agent 3 · green
Email Editor funnel-email-editor Agent 4 · orange
Content content-creator Agent 7 · purple
Social social-media-manager Agent 9 · teal
Designer graphic-designer Agent 8 · pink
Workflow Builder funnel-workflow-builder Agent 10 · purple-2
Web Design Pipeline · Strategist → Designer → Deployer
Strategist page-strategist Agent 11 · yellow
Designer page-designer Agent 12 · pink
Deployer page-deployer Agent 13 · cyan
What I Built

Org structure first. Knowledge library second. Guardrails third.

I built the system the way I'd build a marketing org if I were hiring it from scratch. Three directs report into me: project-manager (Chief of Staff), funnel-strategist (Strategy), funnel-gary-reporter (Reporting). Production agents sit underneath them with explicit scope. The web design team is its own three-stage pipeline (Strategist to Designer to Deployer) because that's how the work actually flows.

  1. 01
    13 scoped agents in .claude/agents/
    Each color-coded, scope-bounded, triggered via natural language. Page-editor (green) never touches checkout. Email-editor (orange) never touches workflow triggers. Page-strategist (yellow) never writes HTML. Constraints are documented in the agent file and enforced at the agent definition layer.
  2. 02
    5-layer shared knowledge library
    agents/knowledge/ stacks WHY Institute foundation → 9 archetypes (per-archetype depth files) → frameworks (CRO, copywriting, email principles, page-design) → audience profiles → agent-infra. Every agent reads the layers relevant to its scope, so brand voice and CRO principles don't have to be re-explained per task.
  3. 03
    Cross-agent state file
    agents/shared/agent-state.json. Every agent reads on start, writes outputs back. Audits land in audits/daily/ and audits/weekly/ (append-only). Recommendations land in recommendations/. Design briefs in pages/briefs/. Builds in pages/builds/<slug>-v<N>/. Historical context never overwrites.
  4. 04
    Two-gate approval workflow
    Bella approves design and copy in ClickUp. Aidan gives final ship in conversation. Comment beats status (fail-closed against "wait" comments). Nothing auto-publishes. Ever.
  5. 05
    Tool integrations across the stack
    Mixpanel (service account), ClickUp (REST + workspace polling for Bella's reviews), Slack (MCP plugin), WordPress (REST + surgical Python patchers, idempotent re-runs, backup snapshots pre-deploy), GHL (REST + Chrome MCP for the workflow builder UI), Google Sheets (gspread service account), fal.ai + Gemini + Canva for the image generation pipeline.
  6. 06
    Versioned, append-only outputs
    Every audit, recommendation, brief, and build is versioned. No silent overwrites. The system stays auditable months later.

The org chart wasn't a metaphor. It was the load-bearing design choice. Scope a real role for each agent, give every agent a place to read from and write to, and the orchestration starts to feel less like prompting and more like delegation.

The Result

Marketing ops, delegable to AI, with the safety rails of a real team.

The daily and weekly audit cadence now runs from a natural-language trigger. Page production goes from request to live in hours with Bella's ClickUp approval and Aidan's ship gate enforced at every deploy. The system has zero auto-publishes by design. Backup snapshots and post-deploy event verification (within 5 minutes) catch problems before they become incidents.

  • 13 agents in production, each scoped to a single function and color-coded to match its org role
  • 5-layer shared knowledge library that every agent reads, so brand voice and CRO principles don't have to be re-explained per task
  • 2-gate approval workflow (Bella in ClickUp + Aidan ship) wired into every web deploy
  • 0 auto-publishes; comment beats status (fail-closed against "wait" comments)
  • Daily and weekly funnel audits replaced (previously hours of multi-system data pulls and spreadsheet stitching) with one natural-language trigger that produces a written report plus ranked recommendations
  • Page production compressed from days to hours, with WP draft + Elementor Canvas + automatic backup snapshot + post-deploy event verification baked in
  • Versioned, append-only outputs across audits/, recommendations/, pages/briefs/, and pages/builds/ so historical context is never lost
What This Proves

Single-person ops scales when the system has a real org chart.

Most AI workflows fall over at the seams. One prompt does too much, the brand voice drifts, the deploy step has no human checkpoint, and the operator ends up doing more cleanup than the agent saved. A bigger model rarely fixes any of that. What fixes it: smaller scope per agent, a shared knowledge layer they all read from, and approval gates wired in as load-bearing structure (not as polite suggestions).

If you wouldn't let a junior contractor publish a page without review, the agent doesn't get to either. The approval gate is the product, not the friction.