Press Workshop
Can a 7-agent crew generate, edit, and publish commercially viable novels? Two books shipped. Here's what we learned.
A single LLM context window breaks at novel length. At 80,000 words, continuity fails and voice drifts — not because the model can’t write, but because it can’t hold 300 pages in context at once. The question was whether a multi-agent pipeline, with each agent owning one stage and reading from disk rather than conversation, could produce fiction good enough to publish and sell on Amazon KDP.
Two books shipped. Read both below.
Winters Bay
Nora Cole became an FBI profiler to understand the kind of man who kills women like her mother. Fifteen years after fleeing Winters Bay, she returns for her mother's funeral — officially an accident, officially a conclusion too convenient to believe.
Mountain Haven
Miranda Castillo inherited a failing diner in a town she left at eighteen. Jake Mercer is the contractor who keeps showing up with tools and grief metaphors. A small-town romance about the things you build when you stop running from what broke you.
How it works
Seven agents, one stage each. Disk-based handoffs — no agent ever holds the full manuscript in context.
Radar surfaces genre trends, pitches concepts. Architect builds the complete structural blueprint before Ghost writes a word: chapter-by-chapter outline, character profiles, series bible. Ghost writes the manuscript in batches of 3–5 chapters, resetting context between batches — the disk-read-from-outline pattern is what makes 80,000 words possible. Mirror runs developmental and line edits file by file. Lens polishes: AI pattern elimination, final grammar. Press packages for KDP — metadata, blurb variants, front/back matter. Shelf maintains continuity state throughout: character tracker, timeline, unresolved threads.
What broke in the POC
Everything She Forgot (~80k words) completed the full pipeline, then went through five publisher rounds. Verdict: “conditional yes — 2–3 passes from ready.” Two failures surfaced.
Handoff failure. Ghost announcing “handed off to Mirror” was text describing intent. Mirror’s session had no access to it. Every pipeline transition required a manual ping — not once, every round, every transition. Fix: sessions_send now delivers handoffs directly to the receiving agent with revision summary and changed files. Manual ping is the fallback, not the mechanism.
Verification gap. Round 4: Mirror and Lens assessed the manuscript as “succeeds brilliantly.” Publisher response: “NOT READY AS-IS.” Of 6 fixes Ghost claimed, the publisher confirmed 0 resolved. Mirror was approving revision notes, not verifying the manuscript. Fix: verification gates now require Mirror to grep the manuscript files for any term Ghost claimed to remove. Claimed fix ≠ confirmed fix. The editorial bar was also recalibrated against senior developmental editor standards — the previous bar was too lenient for commercial publishing.
Results
With those fixes in place:
Mountain Haven — 3 rounds, 6 editorial issues flagged in Round 2, all 6 confirmed resolved via manuscript verification in Round 3. Zero new issues introduced. Pipeline closed its own loop.
Winters Bay — 85,013 words, dual-POV literary thriller, 35 chapters. Mirror Grade A, clean Lens scan. No continuity failures across the full manuscript.
The pipeline works. Whether readers buy the output is the open question, and it’s the only one the market can answer.