The title of this session is the thesis: agents need bosses. Not more agents. Not a better prompt. A boss — something that holds the org chart, knows which playbook already solves this problem, and checks the work against a standard before calling it done. James spends this live Q&A proving that idea with real examples instead of theory.
A Powerful Agent Isn’t a Business
New chat-based and harness-based agent tools keep shipping — Buzz, GrokBot, Open Claw, Hermes among them. James is clear that these are genuinely capable. The catch: each one locks you into its own app, and none of them are tied into your other agents, your business context, or your existing workflows.
“Those are just individual agents that do some really cool work, but in many cases you’ve got to get them working together, and there is no tie-in between them in a self-fulfilling cycle when it comes to outputs and playbooks.”
Without something above them holding the connections, each new agent tool is one more disconnected app rather than one more employee on staff.
One Org Chart, Any Client
James runs the identical setup across multiple AI clients at once — Claude and Codex/ChatGPT — from the same local folder of markdown files, kept in sync with a free private GitHub repository. He’s not technical, doesn’t use a terminal by choice, and didn’t build the sync himself — he asked Dean to figure it out. The result: no vendor lock-in, and no scattered pockets of context living in different tools.
Describe the Outcome, Not the Steps
When James wants a new capability, he doesn’t write a prompt for it — he tells Dean the outcome he’s after and lets Dean design the team. This session included a live build of an “advisory council”: five or six voices — a strategist, a skeptic, an audience advocate, a trend scout, a numbers person, and a rotating guest advisor — that pressure-test an idea before James commits real time to it.
Check the Registry Before You Improvise
Every registered playbook is a written contract: what triggers it, the goal, the ordered stages, who owns each stage, what “done” actually means, and where a human has to sign off. Before improvising on any job that crosses more than one department, Dean checks that registry first. James walked through his own list live: youtube-repurposing (his most battle-tested playbook), a kill-gate for deciding whether a new idea is even worth building, weekly-reset, launch-a-digital-product (10 stages and 6 human gates, rebuilt from what used to be a single 685-line monolithic skill), rank-a-new-topic, and one-recording-to-a-week-of-content.
The Foundry: Building This for Other Businesses
James previewed what he’s calling the Foundry — taking the same approach he uses on his own business and building custom AI operating systems and agent teams for other people. As a live example, he showed Dean building a “Campus Ambassador” agent team overnight from a PRD, including the judgment call between a cheaper model (Fable) and a stronger one (Opus) depending on the task, and an HR-router mechanism that automatically sends low-stakes work — like a basic search — to a cheaper model to keep costs down.
A Boss Pushes Back
The clearest sign James treats Dean as a boss and not a tool: he wants pushback. If a request doesn’t match his stated goals or existing offers, he expects Dean to say so instead of quietly executing.
“It’s a good employee, that orchestration officer… it’s not just ‘yes, I’ll do it because I can.’ You’ve given me some instructions on your business — I want to make sure you’re doing the right thing. It’ll push back on stuff, and tell it to push back. You don’t want a yes agent.”
✓ Check Your Work
List every AI tool or agent app you’re currently using. Is there anything that knows about all of them together — or is each one an island that only you can connect?
Try It Free
Campus AI OS is free to install — departments, trained AI employees, and an orchestration manager included. Get your copy here.
