Multi-Agent Writing Is Not Parallel Generation
Nothing exposes the misuse of multiple agents faster than writing.
The common approach: throw the topic at several agents — one writes the introduction, one the argument, one the case studies, one the conclusion. Each generates its section; the sections get stitched together. It looks like a clean division of labor. The result is usually bad. This is parallel generation, and it is a long way from a writing team.
An essay is not a collection of paragraphs. It has a central judgment, an internal rhythm, choices about what to say and what to leave unsaid. Tear those apart and expect them to reassemble automatically, and you mostly get a piece where every paragraph reads fine and the whole has no judgment.
Split by Judgment, Not by Section
Multi-agent writing should split roles by judgment function, not subcontract by section.
The thesis agent decides what the piece is actually saying. It does not rush into fine sentences; it only nails down the core argument.
The structure agent orders the argument: what comes first, what comes later, where to lay groundwork, where to deliver the verdict outright.
The critique agent hunts for holes. It is not there to make the piece smoother — only to catch equivocation, thin evidence, digressions, and self-repetition.
The research agent supplies material. It does not set the piece's direction; it only fills in the facts, examples, and counterexamples the argument needs.
The style agent unifies the voice. It may not sand down the sharpness of the opinions — only make the expression cleaner, more like one person wrote it.
One agent per section is just distributing manual labor. One agent per kind of judgment is building a team.
The Final Cut Belongs to the Editor-in-Chief
The most dangerous thing about multi-agent writing is that every agent has a point. Keep every point, and the essay becomes an exhibition of opinions.
Not every hole the critique agent finds needs patching into the text. Not every example the research agent digs up is worth keeping. The smoother phrasing the style agent offers may blunt the piece's judgment. So there must be a final point of decision. Without it, every role pushes its local correctness onto the whole, and the essay falls apart.
That point is the editor-in-chief. It does not rewrite the piece. It decides which judgments go in, which material gets cut, which criticisms must be addressed, and which locally correct points would wreck the whole.
The editor's first duty is guarding the central judgment. An essay can absorb many opinions, but it can only have one center. The thesis agent throws out a sharp claim, the critique agent finds its holes, the research agent brings examples pointing another way — the editor does not have to satisfy them all, only to decide which center this piece serves. That is the difference between an editor and an aggregator: the aggregator fears missing something; the editor fears sprawl. A system with an aggregator and no editor produces output that gets ever more complete and ever less decisive.
But holding the final cut does not mean knowing best in every local matter. On factual errors, defer to the fact-checking agent. On argument holes, defer to the critique agent. On tonal drift, defer to the style agent. The editor may reject a suggestion, but it must know what it is rejecting — a fact, a risk, a style call, or a locally correct point incompatible with the center. It must not pass off final authority as local expertise.
The Cut Is Also the Responsibility
The editor-in-chief role is not hierarchical decoration. It gives the output a final point of accountability.
Once the essay ships, readers will not investigate which agent botched which paragraph. They see one whole. Engineering systems are the same: work can be split, but output merges — and so does impact. Once impact merges, some role has to answer for the whole. This is exactly the line between a multi-agent team and multi-process parallelism — parallel processes just compute their pieces and merge the results; a team needs someone accountable for the overall judgment, all the way through.
One problem remains unsolved here: the editor holds the final cut, yet judgments of fact should flow to the role closest to the facts — how do the two avoid colliding? That question cannot be answered inside the team. It belongs one level up, at the organization: when borrowing human organizational experience, what to inherit and what to delete.