Handing a complex project to one all-purpose AI pollutes the conversation and shatters consistency. The moment you split work across role-based specialists with their own contexts, depth changes. Today we unpack the division of labor slowly.
If you've ever tried to write a book with AI, you've probably noticed something strange. The first 100 pages come out fine. Then the main character's personality starts drifting. Settings get fuzzy. Decisions you clearly made yesterday vanish today. Let's walk through why that happens — and how to fix it. Slowly.
First, the principle. Complex work doesn't get done by one person. When you build a house, one worker doesn't draw the plans, lay rebar, run plumbing, and hang wallpaper. You have an architect, a steel worker, a plumber, an interior designer. Why? Because each role uses a different kind of attention. Architecting uses space thinking. Rebar uses number-and-spec thinking. Try to swap between them constantly in one head and mistakes pile up.
When humans work together, this is so obvious no one questions it. But in front of AI, the common sense disappears. We treat AI as one smart generalist. So we pour novels, characters, edits, email drafts, and even 'what should I eat tonight' into the same chat. Identities melt into each other.
As of August 2025, Claude Code shipped a feature called sub-agents. Other AI tools are converging on the same idea, so don't read this as a Claude-only thing. It's an industry-wide shift.
Claude Code lives in the terminal — that black window with only text, what we used to call DOS. You talk to AI there while it handles many files at once. Developers started using it first, but now book authors, researchers, and business planners use it too. Any work where multiple files interlock ends up here.
Inside Claude Code you type one command:
/agents
A screen comes up for creating sub-agents. For a novel project, I might set up a team like this.
Each has its own conversation. When I say 'write chapter 3' in the main window, the orchestrator AI calls book-writer to draft, passes the draft to book-editor to polish, and asks character-keeper for current character state in between. Four agents work in four separate rooms.
The easy way to picture this: a special forces team. In a hostage rescue, no one goes in alone. You have a communications specialist, a medic, a sniper, a demolitions expert. Four people, four missions. Give the medic a sniper rifle and the mission collapses. Each has their own gear, their own training, their own territory. They don't invade each other's lanes.
Sub-agents are exactly that structure.
If I ask 'what should I eat tonight' mid-book, that question stays in the main window only. It doesn't leak into book-writer's context. So chapter 3 doesn't suddenly mention kimchi stew.
The numbers surprise people. I did a 50,000-word book draft two ways.
| Method | Context pollution | Chapters staying consistent | Token usage |
|---|---|---|---|
| Single generalist agent | ~30% | 3 chapters | baseline |
| 4 sub-agents | ~5% | 10+ chapters | 1.4× baseline |
With sub-agents you spend 1.4× more tokens — but consistency jumps more than 3×. You've swapped bottlenecks. Tokens are a money problem. Consistency isn't. Saving a little money while losing the book is worse than spending a little more and saving the book.
Here's the aha.
Pouring everything into one chat is like shoving everyone into the same room and running a loud meeting. Split the rooms and the noise drops.
Where does this pay off most? Ask one question.
'Does this work stretch across many days?'
One-day tasks don't need sub-agents. It's like deploying special forces to pick up snacks. But days, weeks, or months of work is different. Book writing, business plans, research papers, long-running coding projects — for these, sub-agents become a survival skill. One chat window can't carry you that far.
Smallest experiment, for people planning a book.
1. Install Claude Code (Node.js first, then the claude command)
2. cd ~/my-book
3. Type: claude
4. Type: /agents
5. Pick 'Create new agent' at project level
6. Role: "Writes novel prose. Focuses on voice."
7. Tools: read + edit only (no execute — safer)
8. Model: Sonnet (default)
9. Assign a color (yellow makes it easy to see when active)
10. Repeat for editor, keeper, architect
Now ask 'write chapter 3' in the main window. The orchestrator hands off to book-writer. Yellow dot means book-writer is working. Awkward at first. But after ten runs, you understand why people say 'I can't write a book without agents anymore.'
One more thing — model selection.
You can assign a different model to each agent. Put the biggest model (Opus) on the book-writer, a small and fast one (Haiku) on an agent that only cleans up documents. Bolt the biggest model onto every agent and your bill collapses. Slap the smallest one on everything and the prose thins out. Like matching troops to the mission, pair each role's weight with the right model and you balance performance against cost. Almost every AI tool since 2025 supports this combination, so build the habit from day one.
Summary.
Don't use one generalist for complex work. Split by role. Their own context, their own tools, their own lane. In 2025 this concept shows up as 'sub-agents,' but the label will change. It'll get renamed to something else soon. The principle stays: complexity demands division of labor. Humans have known this since the 18th-century factory floor.
One sentence to remember — 'One job, one agent, one conversation.' Hold that rhythm and your AI collaborations run much further. Three-month projects reach a year. Books that stalled at chapter 1 become actual books. The person who assembled the team completes the mission. The person clinging to one generalist burns out in the middle.
Technology changes. Names change. Principles don't.
Divide. Delegate. Combine.