Building a multi-agent SEO system, the architecture behind OpenClaw

Most "AI SEO" tools today are a single prompt with a fancy UI. You paste in a keyword, the LLM returns an outline, and you call it a content brief. It looks impressive in a demo. It collapses the second you try to run it across 10 client tenants.

OpenClaw is the architecture I wish existed when we started running SEO retainers. This post is the why and the how.

The single-call problem

When you wire up an SEO workflow as one LLM call, you bake every responsibility into one prompt:

Keyword research happens in the same context as brief drafting
Internal-link suggestions get blurred with content angle
Reporting is a separate manual step that lives nowhere

That’s fine for one client. At ten clients, the prompt becomes ungovernable. Different clients need different defaults. Brief structure varies by industry. Internal-linking strategy depends on the existing site graph. The single-call approach scales the work linearly with client count, you might as well still be doing it manually.

Four agents, one orchestrator

OpenClaw splits the work across four specialized agents, each with a narrow job:

Keyword agent

Pulls real search-intent data for a client niche, clusters terms by topical group, ranks them by opportunity (volume × difficulty × commercial intent). Output: a structured keyword graph the brief agent can consume.

Content brief agent

Takes a keyword cluster and produces a production-ready brief, angle, sources, internal-link targets, target term, suggested H2 structure. Optional auto-draft if the client wants a first pass instead of a brief.

Internal linking agent

Crawls the client site, builds a graph of existing pages, identifies orphaned content and link-equity gaps, suggests (or applies) internal links that actually move rankings.

Reporting agent

Pulls live data from each integrated source (Search Console, GBP, GA4, the client’s CRM if connected), assembles a weekly client report under the agency’s brand, and ships it.

The orchestrator routes work between them, it knows that the keyword agent’s output feeds the brief agent, and that the brief agent’s "internal_links" field gets validated by the internal-linking agent against the actual site graph.

Why this architecture wins

Three things you can do with split agents that you can’t do with a single call:

1. Specialize per agent

The keyword agent runs against a model that’s good at structured output (we default to GPT-4o for this). The brief agent runs against a model that’s good at long-form (Claude Sonnet 4.6 is our default). The reporting agent doesn’t need an LLM at all most of the time, it’s pulling structured data and rendering a template.

A single-call architecture has to compromise on the model choice for whichever step is most expensive.

2. Reuse work across clients

Keyword research for one HVAC client in Miami informs keyword research for another HVAC client in Tampa. The orchestrator caches term-level data and reuses it across tenants when geography permits. A single-call architecture re-does the work every time.

3. Multi-tenant isolation

Each client has their own configuration, niche, target keywords, brand voice, integration credentials. The orchestrator routes each agent invocation against the right tenant context. A single-call architecture has to either re-prompt the entire context each time (expensive) or rely on fragile global config (error-prone).

What we punted

A few things that look obvious but we deliberately didn’t build:

A single-pane-of-glass UI. OpenClaw is API-first. Reports go to Slack or email. Briefs go to the client’s existing CMS or doc platform. Building yet another dashboard nobody opens isn’t leverage.
Auto-publishing. The brief agent ships briefs to your writer, not WordPress. Auto-published AI content is how you get penalized.
Real-time everything. Most of the work is batch. Keyword research runs weekly. Briefs run on demand. Reports run weekly. There’s no value in running this synchronously and a lot of cost.

What’s next

OpenClaw is in early access through 2026. Self-hosted is free; managed pricing announced at launch. Waitlist members get 50% off year one.

If you’re running SEO for more than 5 clients and tired of duplicating manual research, join the waitlist. If you want to talk through the architecture before signing up, book a discovery call, happy to walk through the orchestrator details.

More posts coming as we ship: model selection per agent, the cost economics, how the multi-tenant control plane handles auth, and what we learned from the first cohort.