AI SEO agents built in Claude Code

An AI SEO agent does the research, not the deciding

The agents in these builds do one thing well: research that can run in parallel. Each AI SEO agent is a Claude Code subagent with a narrow job, a limited set of tools and a fixed format to report back in. One orchestrating session reads what they return, checks it, and is the only thing allowed to write to the shared files.

They run across three kinds of work so far: long-form articles, a software directory and a directory of food distributors.

Why one big agent goes wrong

Asking a single agent to research, write and publish sounds efficient. In practice it writes over its own files, trusts its first search result, and produces output nobody can check piece by piece.

Splitting the work fixes most of that. Research agents run side by side and return a pack, a JSON record or a dossier. The orchestrator, which holds the rules and the context, decides what survives. That split is the same reason most useful AI agents for SEO end up looking less like a writer and more like a research team.

The agents and what each one may touch

Each agent's tools are chosen for its job, and the list is part of the design:

The rules are enforced by those tool lists plus written instructions. The article researcher can run shell commands for curl, so "it can't write" is an instruction it follows, not a wall it can't climb, and the orchestrator is written accordingly.

Research for articles

For each article in a batch, one researcher runs in parallel with the others. It reuses facts from the client's knowledge base if they were verified within about 60 days, pulls 15 search results from Ahrefs and 10 from Jina, keeps and scrapes the 10 most relevant, and checks prices live.

It also returns two Reddit insights and two authoritative research sources, and hands the pack back. A separate capturer takes screenshots and has to open each image to confirm it shows the real page, not a login wall. The writing itself stays with content creation automation, where the drafts go through five proofreading passes.

Research for a directory

On a software directory, one researcher per market segment returns up to 18 tools each, with exactly four features per tool. A profiler agent writes each tool's profile as JSON, and a loader script refuses anything thin before it reaches the site:

• Fewer than four strengths, two limitations or three sources
• An overview under 90 words
• A missing scorecard dimension, or an em dash anywhere

On the VoiceOrder distributor directory, up to five market researchers run in the background, each writing a dossier and saving the pages it captured. Pages are then built one at a time from those dossiers, with at least 10 verified suppliers on each. The directory now has 43 market pages across 13 cities, with 789 verified listings.

What the orchestrator catches

Research agents are wrong sometimes, and the design assumes they will be. In one batch a research pack moved the timeframe of a published statistic, another reported a competitor's figures as changed when they hadn't, and a third described content on a pricing URL that only redirects to a homepage. The orchestrator caught all three by opening the pages itself.

Other failures shaped the rules. A clash over a shared file in the first article batch is why only the orchestrator writes shared files, and on the directory two research agents created the same file and one overwrote the other. When a dossier's note to leave a phone number blank was read as the phone number, the quality script started checking the shape of every value.

What isn't automated yet

Writing, proofreading and SEO agents were planned for the article pipeline and not built. Those stages run with a person driving them, and for now that is deliberate: they are where judgement matters most. Nothing in these builds re-checks a directory listing after it is first verified either, so freshness is a scheduled human job.

When a team of agents fits

It fits research-heavy work that repeats: batches of articles, directories with dozens of entries, or market pages that each need verified facts. The more pages share a research pattern, the more an SEO agent setup pays back.

It doesn't fit a one-off article, or work where the research is really an interview. Agents are good at gathering what's already public, and no use at all for the part only your team knows.

What a research agent costs

The agents run inside Claude Code, and the research tools they call bill per request, so a batch costs little compared with the time it replaces. The setup is writing each agent's instructions, tools and return format.

The agents come with execution capacity: $2,000 a month flat for the AI agents route, taken with a fractional Head of Content retainer. A new research agent for an existing pipeline is a matter of days; a new pipeline is closer to two weeks.

Start with the research you repeat

List the research your team does again and again, like pricing checks, competitor facts or supplier lists. On a call we'll pick the one that suits an agent best and sketch what it would return.