What You’re Going To Learn

If you were starting your AI agent team from scratch today, which roles would you hire first? That’s the question James answers in this live session — and he backs it up by showing the actual system running his business right now.

You’ll see the orchestrator (Dean), the four department managers, the three memory layers, and how scheduled tasks run the whole operation overnight while James sleeps. Plus a live walkthrough of the outbound sales pipeline and the YouTube repurposing waterfall.

Step 1 — Start With an Orchestrator, Not Individual Skills

Most people build AI tools the wrong way. They install a skill here, add a plugin there, and then they become the bottleneck — manually deciding which tool to call, in what order, every single time.

The fix is an orchestrator. James calls his Dean. Dean knows every department, every employee, every tool available. When James says “I want 10 more leads” or “publish this video,” Dean figures out which agents are needed and routes the work — without James having to manage the steps.

Think of it like a waterfall with locks. Without an orchestrator, there’s one gate — everything moves in sequence, one task at a time. With an orchestrator who can open five gates at once, work flows in parallel. That’s the difference between a task queue and a real team.

The first hire isn’t a skill. It’s the person who manages all the skills.

Step 2 — Organize Into Departments, Not a Flat List of Skills

Once you have an orchestrator, the next decision is structure. James broke his team into four departments:

• Marketing Manager — YouTube, social, newsletter, content repurposing
• Sales Manager — outbound pipeline, CRM, campaigns, launches
• Education Manager — course structure, live sessions, lesson creation
• Community Manager — campus engagement, onboarding, daily posts

Each department manager has its own rules, memories, and workflow recipes. When James tells Dean “do the social media for the week,” Dean hands off to the Marketing Manager — who already knows which employees to call, what tools to use, and in what order.

You can structure departments by function (like James), by client, or by project. The structure doesn’t matter as much as having one. Without it, you’re still the one deciding which skill to call next.

Step 3 — Build Three Layers of Memory Before You Scale

Agents without memory make the same mistakes over and over. They forget your voice, your customers, your goals. Memory is what turns a smart tool into an actual employee.

There are three layers every AI operating system needs:

Layer 1 — The Static Filing Cabinet. This is the bedrock: ideal customer profile, what you sell, your business goals, what tools are available, who the business serves. It gets set once and re-evaluated regularly. Every agent reads this before doing any work.

Layer 2 — Living Memory (memory.md). Every time an agent completes a task, it logs what worked, what didn’t, and what to remember next time. This is the running operational record — what’s in flight, what was just shipped, what needs follow-up. Claude reads this at the start of every session so it never starts cold.

Layer 3 — The Wiki (indexed knowledge base). This is the long-term library. Everything that compounds over time — decisions made, patterns discovered, strategies proven — lives here. Inspired by Andrej Karpathy’s LLM Wiki concept. The more context you accumulate, the better every future output gets.

Whoever has the most relevant context wins. Google IO made this obvious: every demo that wowed the crowd only worked because Google already had the user’s data. Your campus and your personal library is your version of that advantage.

Step 4 — The First Two Agent Pipelines Worth Building

Once the orchestrator, departments, and memory layers are in place, these are the two pipelines James built first — and recommends for anyone starting out:

Outbound Sales Pipeline

Five employees, one sequential workflow:

1. Scout — searches LinkedIn, YouTube, Reddit, directories for prospects matching your ICP
2. Enrich — researches each prospect: what they sell, who to, recent launches, pain points (20–30 min per person, done automatically)
3. Score — rates each prospect on niche fit, intent, timing, and accessibility
4. Brief — writes a one-page research memo per top prospect
5. Draft — writes personalized outreach (cold email, DM, follow-up) ready for your review before anything sends

Result: 15–25 researched, scored, drafted prospects per week without you touching any of it. Human stays in the loop at every approval checkpoint — or removes themselves once they trust the process.

YouTube Repurposing Waterfall

One YouTube video in. Five finished pieces of content out.

1. Transcript pulled automatically via YouTube Transcript connector
2. Content brief created — key messages, quotable moments, structure extracted
3. Tutorial written and published to WordPress (SEO-optimized, YouTube video embedded, live immediately)
4. Email announcement drafted in FluentCRM — points to the tutorial, not YouTube
5. Community post published to campus
6. Social posts fired to LinkedIn, X, Threads, Facebook, Bluesky via Blotato

James now has this running as a scheduled task. A new video goes up, and 24 hours later the full waterfall fires automatically. He doesn’t open the admin. He doesn’t write the email. He just makes the video.

Step 5 — Connect the Right Tools

Workflows without connectors give you documents. Workflows with connectors get things done.

Without a CRM connector: your sales agent produces a spreadsheet. With it: prospects are added directly, social profiles attached, research notes filed, outreach queued.

Without a publishing connector: your tutorial agent gives you text. With it: the post is live on your site in the right format, with taxonomy, author, featured image, and all meta fields populated.

James uses WordPress as his central data hub — FluentCRM for email, FluentCommunity for campus, FluentCart for products, all under one login. But the Campus AI OS works with Kajabi, GoHighLevel, ActiveCampaign, and others — if a connector exists, the agents can use it. If not, they produce the output for you to place manually.

Do a tool inventory. Check what connectors are available for each platform. That gap between “gives you a doc” and “does it automatically” is almost always a missing connector.

Common Mistake

Starting with skills before you have a workflow. Installing a dozen skills and having no orchestrator to call them means you’re still doing the coordination — just with fancier tools. The correct order: orchestrator first → department structure → memory layers → individual skill pipelines. Build the scaffolding before you hang the art.

What to Do Right Now

Pick one workflow you do manually every week. Map the steps. Identify the employees you’d need. Then start with that one pipeline — the orchestrator, one department, one workflow. Get that running before you add more. Complexity compounds, but so does the time savings.

The Campus AI OS plugin installs the orchestrator + four departments + 20 employees in one step. The free Campus Agent OS on GitHub gives you the structure without the employees. Either way, this is available now — not in six months.