SEO is three jobs. Only one of them is writing.

People say they want an SEO agent. What they mean is they want traffic without picking a fight with three different jobs.

Research wants sources, gaps, and pages that already rank. Writing wants a draft that sounds like you. Review wants a cold checklist and no new ideas. Put those in one head and the writer grades its own homework. Of course it passes.

One SEO agent is the wrong unit. Three jobs. The third one is why you do not publish a confident mess.

Research is a fetch job

Scout opens the live pages you named. Competitor URLs. Your last three posts on the topic. The doc where you already decided the term. It comes back with quotes, headings, and a one-page brief: what they cover, what they skip, what we can say because it is true.

It does not invent search volume. It does not "know the SERP" from 2019. If the number is not on a page it can load, it leaves the cell blank. Blank is honest. A made-up difficulty score is how you spend a week on a word nobody types.

Pulling pages is a job with a computer. It is not a magic index.

Ahrefs and Semrush still win on the historical link graph. Pay them for that. Do not ask a teammate to hallucinate a backlink profile. Different sport.

You still pick what to rank for. That is taste plus a spreadsheet. First-time invention. Keep it.

Writing is a brief job

Maker gets the brief, the voice file, and the offer. Not "write something about invoicing." The brief says the term, the reader, the three points, the one claim you will stand behind, the internal links, the length.

Then it writes. Then it stops.

If Maker also picked the term, you will get the term that was easiest to waffle about. If Maker also reviews, you will get a draft that "flows." Flows is how 2,000 words of nothing gets a title.

The artifact is a draft in Needs you. Not a URL. Publishing is still a human with a CMS and a stomach.

Review is the job that makes it safe

"Quality" is not checkable. People fight about quality. A checklist is not a fight.

Did it hit the word band. Did it use the term in the title and the first screen. Did it avoid the banned phrases. Did every claim trace to a source in the brief. Is the one comparison table actually a table. Did it link the three pages you named, and only those.

Watch can run that list. Or a third teammate whose only personality is "no." Same model is fine. Different job, different memory, different stop rule. The reviewer does not get to rewrite. It scores. Fail goes back to Maker with the line that failed.

That is why review is its own lane. Rule-compliance is an artifact. Vibes are not.

Would this force a context switch. Yes, if the writer has to become the cop. Split it.

Can they fetch without inventing. Yes, if the brief named the URLs. No, if you asked for "what Google wants this month" with no sources.

Do the rules of job A poison job B. Writing wants momentum. Review wants friction. Mixing them is how friction loses.

What a Tuesday looks like

You already chose eight terms for the quarter. That meeting happened. It will not happen inside the model.

Monday: Scout returns briefs for two of them, each with live pages attached.

Tuesday: Maker writes the first. Parks it.

Wednesday: the reviewer scores it. Two fails. Maker fixes those two. You read the thing once, because the cheap mistakes are already gone.

Thursday: Mail cuts a note for the list. Different lane. Marketing splits the same way. SEO does not get to write the subject line.

Nothing hits the site without you. Caps on sends. Keys in the vault. Dedicated workspace. We never train on the draft.

Honesty about the ceiling

This will not recover a domain you wrecked with spun pages in 2012. This will not place a link on a newspaper. This will not tell you the algorithm.

It will turn a list you already believe in into drafts that obey a brief, on a schedule, with a log. That is the job that recurs.

LangChain's numbers had research and data around 24.4% of agent deployments. SEO research is that job with a URL list. Treat it like research. Do not treat it like an oracle.

$99 on Octopus is OSS models, a computer, and a board. Bring ChatGPT, Claude, or Grok if you already pay. No markup. Seven-day money-back if the three-job split feels like babysitting. For most teams it feels like finally knowing who to yell at.

Hire Scout for the pages, Maker for the draft, a reviewer for the list. Chief routes so the brief does not die in a chat. You keep the terms. The link tools keep the graph. Everybody stays in their lane, which is the whole trick.