Video: "Hermes + Claude Agent Swarms AI SEO is INSANE!" by Julian Goldie on YouTube.
What changes when Claude drives the swarm
Hermes Agent has had swarm mode for a while — you can split work across multiple agents and let them run in parallel. What's newer is running Claude as the model behind each agent in the swarm, rather than a local open-source model. Claude's ability to hold context, follow structured instructions, and produce editorial-quality output means the individual agents in the swarm are noticeably more reliable. The planner stays on brief. The reviewer actually catches problems rather than just rubber-stamping the previous step.
Worth noting: the difference between a swarm where each agent uses a capable model and one where they use a cheaper local model is substantial. The quality of the routing logic — which agent gets which task, and when — depends directly on the model doing the routing. Claude handles that step well.
How the SEO swarm runs in practice
The setup Julian demonstrates works like this. You give the swarm a target: a keyword cluster, a topic area, or a specific content goal. The router agent breaks that into discrete tasks and assigns them. The planner agent builds the keyword strategy, topical map, and content structure. The builder agent writes the actual content to that structure. The reviewer agent checks the output, flags gaps, and suggests edits before anything is finalised.
All of this runs in parallel where tasks don't depend on each other, and sequentially where they do. The practical result is that a workflow which would take a human content team half a day — brief to draft to reviewed output — finishes in minutes. The keyword research, topical map, internal linking logic, and multiple content outlines come out together.
What the output actually looks like
The swarm produces structured markdown: keyword clusters with rationale, page structures for each topic, suggested internal link anchors, and draft copy for each page. The quality is consistent across all of it, which is the real advantage over prompting a single model repeatedly. When you prompt one model for each task in sequence, the context gets thinner as you go. The swarm maintains a shared brief and each agent draws from it, so the content on page seven doesn't drift from the logic set on page one.
That said, the outputs are drafts. The keyword clusters need checking against real search volume data. The copy needs someone who understands the business to read it. Hermes doesn't know your actual customers, your margins, or the angle that makes your service different from a competitor. That context still has to come from you, in the brief you write before the swarm starts.
Where the limitations are
Two areas need watching. First, the swarm can only act on the brief it's given. If your brief is vague, you get plausible-sounding output that doesn't quite fit. The old rule applies: better in, better out. Second, the reviewer agent is not an SEO expert. It checks for consistency and coherence, not for whether the page will actually rank. Running the output through a proper SEO check — search intent alignment, entity coverage, heading structure — is still a manual step.
In practice, the swarm is best used to get to a solid first draft quickly, not to produce finished content without oversight. Most teams will find it compresses two or three days of content planning work into an hour, then they spend the rest of their time editing and publishing rather than writing from scratch.
Where this connects to NordSys
Setting up a working Hermes swarm with Claude as the model requires getting the agent configuration, skill library, and API keys in the right shape — and keeping that working as Hermes updates. That's exactly the kind of install and ongoing care we handle for clients. If you want a multi-agent SEO workflow like this running for your business without managing the technical side yourself, our AI Agents service covers the full setup, hosting, and maintenance.
See our AI Agents service →