The least interesting use of agents is asking them to write the final thing.
They can do it. Sometimes they can do it well enough. But if the workflow stops there, the agent becomes a fast producer of plausible prose. That is useful in the same way a very confident intern is useful: you get momentum, but you also get a new review burden.
The better use is to treat agents as editors, reviewers, validators, and workflow operators.
Writing is only one part of publishing. A piece also needs source gathering, structure, duplicate checks, title alternatives, humanizer checks, claim checks, metadata, Ghost formatting, tracker updates, Obsidian mirrors, and post-publication checks. Agents are often more valuable in those surrounding jobs than in the first draft itself.
The reason is simple. Agents are good at carrying checklists across messy folders. They are good at comparing versions. They are good at noticing missing tags, broken source labels, stale statuses, repeated headings, and claims that have no obvious support. They are good at turning a loose workflow into a repeatable path.
They are less good at knowing whether a sentence sounds like me.
That is where taste stays human. I can ask an agent to score a draft, suggest alternatives, or find AI-ish phrasing. I should not let it flatten the voice into generic smoothness. The point of a humanizer pass is not to make the piece blandly human. It is to find the parts that sound inflated or mechanical.
The best agent review loops have clear roles.
One role is the drafter. It turns a source pack or outline into a first coherent version. This is useful because blank pages are expensive.
Another role is the critic. It asks whether the argument is actually supported, whether the title overpromises, whether the structure repeats itself, and whether the piece has a reason to exist.
Another role is the validator. It checks the mechanical things humans are bad at checking repeatedly: frontmatter, tag conventions, unfinished scaffold text, word counts, links, source requirements, title formats, index updates.
Another role is the operator. It moves files, updates trackers, mirrors folders, creates Ghost drafts, and verifies that the public artifact matches the local state.
Those roles should not be collapsed into one vague instruction to "make it better." When agents have jobs, they are useful. When they have vibes, they improvise.
This also changes the trust model. I do not need to trust the agent's taste in order to trust its checklist. I can inspect the output of a validator. I can look at a claim ledger. I can compare a Ghost draft. I can read a humanizer score as a signal, not a verdict.
The workflow is strongest when every agent action leaves a trail. What file did it touch? What check passed? What failed? What URL was created? What next step remains? That trail is what makes delegation possible without losing the plot.
The trap is to confuse automation with judgment. A piece can pass validation and still be boring. A title can be accurate and still feel dead. A draft can have sources and still miss the real point. The system needs human taste because taste is the thing deciding what is worth saying.
But taste should not have to spend its energy on clerical work. That is the bargain.
Let the agents do the repeatable inspection. Let them make the boring state legible. Let them propose options. Let them catch the failure modes that happen when a workflow has many handoffs. Then use human judgment on the decisions that actually matter: the angle, the voice, and whether the piece deserves to exist.
This is why I prefer agents as editors to agents as ghostwriters. A ghostwriter tries to disappear into the final prose. An editor improves the work while leaving the author responsible for the final shape.
That responsibility is the whole point. The public notebook should not sound like it was generated by a content machine. It should sound like a person using machines to think and publish more deliberately.
Agents multiply effort. They are not taste. The workflow works when that distinction stays visible.
The practical test is simple. If an agent helps me see the work more clearly, it is doing its job. If it helps me avoid a broken citation, a stale status, a weak title, or an unsupported claim, it is doing its job. If it makes the prose smoother while making the idea less mine, it has crossed the wrong line.
That line is where the editor frame matters. I want help that sharpens judgment, not help that hides the need for judgment.
This is part 6 of 8 in Operating a Public Notebook.