What Is Multi-Agent Writing? How AI Agents Write Novels
Multi-agent writing splits a novel among AI roles: a lead assistant plans with you, a writer agent writes, an archivist keeps the archives. You decide.
In this article 6 sections
What Is Multi-Agent Writing
Multi-agent writing splits the job of writing a novel among several AI roles that each handle one thing: one talks things over with the author and arranges tasks, one writes only the body text, one records the story so far, and one only picks out problems. They hand material to each other, and in the end the author decides whether to use it and how to change it.
Compared with “one chat box from start to finish”, the advantage of dividing the work is that each role gets only the material it needs and is responsible only for its own job.
Why Writing a Long Novel in One Chat Box Is Hard
Ask a general chat model directly to “help me write a chapter of xuanhuan fantasy”, and the result is often quite good. But after several dozen chapters, the problems become more and more obvious:
- The story so far isn’t fully remembered: a character’s injuries, favors owed and foreshadowing planted are more easily pushed out of the context the longer the chat goes on.
- Requirements get diluted: the style and taboos set at the start are followed less and less as the chat continues.
- Writing and reviewing at once: right after the same model finishes a passage, asking it to find faults in its own work often turns up nothing.
These cannot be fully solved just by switching to a stronger model; they are more a matter of how the work is done.
The Division of Labor in Maliang Writer
In the new-generation writing engine of Maliang Writer, the common roles are:
- Lead assistant: the one you talk to. It understands your requirements, reads the earlier text, outline and archives, plans chapter outlines, and decides what to do next and who does it.
- Writer agent: writes only the body text. What it gets is the earlier text this chapter needs, the chapter outline, the relevant characters and style references, without a long chat history, so it can concentrate on writing this chapter well. The model that writes the body text can be chosen separately from the lead assistant’s.
- Archivist: after a chapter is written, organizes this chapter’s summary, characters who appear, character states and relationship changes, foreshadowing and in-story time from the body text. What the text doesn’t say isn’t filled in; anything that conflicts with existing settings is flagged for you to confirm.
- Research and review: when needed, the lead assistant assigns a separate task to look up material, or to run an independent review. A review lists the specific original sentences, locations and suggested changes, and does not edit the draft for you.
The division changes with the task: when you are only talking plot, perhaps only the lead assistant is working; when continuing a chapter, the writer agent and archivist join in. It is not a fixed set of agents taking turns every time.
How They Work Together on a Chapter
Take “continue with chapter 23” as an example:
- The lead assistant reads the last few chapters of body text, this volume’s outline and the current archives of the relevant characters, and confirms with you what happens in this chapter;
- Once the chapter outline is settled, it is handed to the writer agent to write the body text;
- The body text is written into the editor, and you can see paragraph by paragraph what changed, and keep or restore each change;
- After the chapter is written, the archivist updates the records of character states, relationships and foreshadowing, and lists anything in doubt as pending confirmation;
- When you want a second opinion, you can have the lead assistant assign an independent review.
Throughout the process, where the story goes, which passages stay and which are rewritten are all up to you. For more on continuing, see AI Novel Continuation, and for how archives are kept, see Story Archives.
Why Dividing the Work Helps a Long Novel
- Cleaner material for the writer agent: it carries only the earlier text and settings this chapter needs, its style references are more focused, and it isn’t led astray by scattered requirements from dozens of rounds of chat.
- Someone dedicated to the story so far: the archivist records only from the body text, and the archives are read before continuing, so earlier events don’t have to rely on the model “remembering”.
- Writing and reviewing are separate: an independent review isn’t affected by the inertia of “just having written it”, and finds problems more easily.
- The author can step in at any time: every step’s output is visible and editable, and when AI changes the body text you can undo that round.
Dividing the work cannot guarantee zero contradictions and cannot replace the author’s judgment, but it gives each of the steps most prone to error in long-form writing someone to look after it.
FAQ
Q: How is multi-agent writing different from ordinary AI writing?
Ordinary AI writing is usually one model handling everything in one conversation; multi-agent writing gives discussion, body text, recording the story so far and review to different roles, each getting only the material it needs and doing only one thing.
Q: Will the agents take it upon themselves to change my draft?
When AI changes a chapter you have open, the editor shows a before-and-after comparison, and you can keep or restore each change; if you are unhappy with a whole round, you can undo the current round. A review task only offers suggestions and doesn’t edit the draft directly.
Q: Does multi-agent writing use more credits?
The extra cost is mainly archive organizing and review and research tasks as needed. Writing the body text is still the main expense; each round’s actual cost can be viewed in the product.
Q: Can I use only part of it?
Yes. Only talking plot, only planning an outline or only having AI review a chapter are all fine; archive organizing can also be turned off in settings.