Multi-Agent Book Writing

Why one context can’t hold eleven jobs

Writing a long, research-backed non-fiction book alone with Claude works fine until it doesn’t — citations start appearing that no paper supports, voice flattens to generic therapy-speak by chapter eight, and an argument built in the opening chapters quietly splits into two incompatible frames by the midpoint. The problem isn’t prompt quality. Academic citation discipline, narrative judgment, and sensitivity review are different skills with different reflexes that actively interfere when run in one context at the same time.

This book builds the production system that solved those problems on a real 25-chapter academic project, coordinated across eleven specialist agents — each one owns a domain, and none of them step on each other.

Eleven chapters, one roster

• designing-the-roster — Eleven specialist roles, what each owns, what it doesn’t, and the split between the portable core stack and project-specific domain specialists.
• prompting-specialists-not-generalists — The difference between a role description and operating criteria — and why an agent needs both to form a real opinion.
• shared-context-the-coordination-layer — The single document every agent reads first: locked terms, locked structural decisions, and the state that keeps eleven agents from drifting apart.
• running-research-agents-in-parallel — Partitioning a brief across domain research agents without producing claims that conflict at the seam between chapters.
• the-citation-review-agent — A verification protocol that catches plausible generalizations, superseded claims, and directional rounding before they reach the manuscript.
• the-structure-agent — Tracking whether a 25-chapter argument still accumulates toward its central claim, from a chapter list and one-sentence summaries alone.
• the-copyeditor-agent — Register consistency and AI-tell removal at book scale, without stepping into argument content or citation accuracy.
• sensitivity-and-philosophy-agents — The two roles every writer understaffs until the manuscript is nearly done, then discovers they can’t skip.
• running-the-council — Dispatching a draft to four agents in parallel, then synthesizing conflicting feedback into one structured, actionable log.
• when-the-council-sends-you-back — The revision routing protocol that keeps a structural fix from cascading into a second, bigger fix.
• from-manuscript-to-press — Why python-docx beats pandoc for a press-ready file, and the academic polisher’s narrow, late-stage pass before the final build.

A preview

The citation looked right. That was the problem. … You do not catch your own fabrications by reading the draft again. Going back through it is how you launder them into facts.

Writers scaling past what one agent can hold

Writers already running multi-agent workflows in Claude Code who are taking on a long, research-backed, high-stakes manuscript — the kind where a single generalist agent can’t hold citation accuracy, structural arc, and register consistency at once. This is not an introduction to what an agent is; it’s the production system for when the solo-agent approach breaks down.

A longer excerpt is available to newsletter subscribers.

Buy Ebook on Amazon