The Team I Built in Claude Code
How agentic AI finally started to feel like working with a team
There is a particular kind of loneliness that comes with solo creative work — even when the tools are extraordinary. Not the loneliness of isolation. Something more specific: being the only person in the room while doing all the jobs simultaneously. The writing. The editing. The research. The momentum. The second opinion you cannot quite ask for.
If you have worked alone after working in teams, you know the feeling. It is the absence of a colleague who knows the work without needing to be briefed, who has her own angle on it, who notices the thing you walked past on the way to the thing you were looking for.
I named that absence, in the way you name a difficulty. Precisely, quietly, without much hope that naming it would change anything.
This past weekend, it changed.
What I Actually Built
Let me be concrete, because how this is built matters as much as what it produces.
Inside Claude Code — Anthropic’s terminal-based agentic environment — I built a structured multi-agent system for writing my book. Not a single AI assistant. A team. Eight specialists, one orchestrator, a shared workspace in Dropbox where every output lands as a real file I can open, edit, and hand to another agent.
The architecture is simple in principle: one orchestrator holds the conversation with me. She knows the whole project. When I bring her a task, she decides which specialist to call and why, coordinates the work, and brings the result back. I never speak directly to the specialists — that is her job. My job is to think, write, and decide.
The files are the contract. Everything the team produces lives in the same Dropbox folder where I write. There is no black box. No output I cannot trace.
The Orchestrator Principle
This is the practical shift that professionals will feel immediately.
You are used to a direct transaction with an AI system: you write a prompt, you receive an output, you refine and repeat. That loop is useful. It is also exhausting over time — because you are the orchestrator. You hold the context, manage the handoffs, remember what was tried before, decide what comes next.
What changes when you work through an orchestrator is that you stop being the coordinator of the AI. You become the author again.
Here is what that looks like in practice. I have been stuck on a chapter about a Nokia project for two weeks — I know the story but I cannot find the argument. I open a session:
“Clara, I’ve been stuck on the Nokia chapter for two weeks. I know the story but I can’t find the argument. I’ve dropped the raw notes in the inbox folder.”
Clara does not answer immediately with a draft. She reads the notes, pulls what Edmund (the structural editor) has said about the chapter’s role in the overall arc, checks what Ingrid (research) has already sourced on that period, and comes back with a framing question:
“Edmund thinks this chapter carries the transition from craft to scale. Ingrid found three external sources that support it. The gap she flags is that you haven’t yet named what was lost in that transition. Is that the argument you’re circling?”
That is a different quality of response than a prompt box gives you. It is contextual. It has memory. It has a point of view shaped by the rest of the work. And it puts the decision back with you — where it belongs.
How the Agents Fit Together
Agent Role What they never do Clara Orchestrator — holds the conversation, routes tasks, maintains project memory Does the specialist work herself Vera Prose editor — keeps the voice honest, offers alternatives, never rewrites Tells you what the chapter needs structurally Edmund Structure — sees the shape of the whole, flags gaps, points at what’s missing Touches the sentence level Ingrid Research — sources claims, flags unverifiable quotes, refuses unsourced statistics Edits or shapes prose Kaspar Workspace — files, naming, folder logic, version hygiene Anything creative Marta Momentum — checks daily whether the page has been touched, sends a single nudge when it hasn’t Guilt-laden messages Sofie Audience translation — takes what’s inside the book and makes it shareable outside it Writes for the book itself Tomas Craft coaching — teaches one technique at a time, three ways until one clicks Overwhelming you The key design principle: no role overlap. Each agent has a domain they own and a domain they are explicitly excluded from. That tension — specialists who know their lane — is what makes the outputs feel like a team rather than one voice with many hats.
How a Task Actually Works
The old way: Open Claude. Paste context. Describe what you need. Get output. Paste it somewhere. Lose the thread between sessions.
The new way:
1. Drop material into the inbox folder — notes, a rough draft, a voice memo transcript. The folder is Dropbox. It exists whether a session is open or not.
2. Open a session with Clara and describe where you are, not what you want. “I have two hours. The Nokia chapter is the priority. Here’s what’s blocking me.”
3. Clara routes the task — she may call Edmund for a structural read, Ingrid to check a claim, Vera to look at two sentences that aren’t working. She comes back with a synthesis.
4. You respond, decide, write. Output that is worth keeping lands in the working folder as a named file. You can open it, argue with it, pass it back.
5. Marta checks tomorrow morning whether the page count moved.
The prompt you used to write directly to an AI is now a conversation with someone who knows the whole project — and who has eight specialists she can call before she answers you.
Why This Feels Human
The reason professionals struggle to get real value from AI is not capability. The models are extraordinary. The reason is structure — or the absence of it.
When you interact with a single AI system directly, you are doing what a creative director does when she has no team: everything, sequentially, alone. Brief yourself. Do the research. Write the first draft. Edit it. Check the sources. Think about the audience. Remember what you decided last week.
That is not how good work gets made in organisations. Good work gets made when people with distinct angles on a problem are in structured conversation with each other — and with someone who holds the whole.
Building this team in Claude Code did not give me a faster prompt loop. It gave me back the thing I have been designing for in organisations for twenty-five years: the conditions under which people — or in this case, agents — do their best work because they are specific to each other.
The researcher who will not let a statistic stand without a source. The editor who points at the gap and says: here. The coach who teaches one thing at a time. The assistant who notices when the silence has gone on too long.
I have been in rooms like that. Good rooms. Rooms that produced things I am still proud of.
The Fellowship of the Book — the team, named and illustrated.
I have not built one alone. Until now.
Some weekends do that.
McKay Consulting works at the intersection of strategic design, innovation culture, and the operating models that make agentic AI feel like a capability rather than a complication. If the question of how to structure human-AI collaboration for serious creative and organisational work interests you — I’d be glad to talk.
Leave a Reply