A keyword agent that mines and prioritises, a content agent that scores and fixes its own drafts, a backlink agent that screens links and scouts prospects — all sharing one memory. Below: one cycle on our own site, produced with Claude for this proposal.
This cycle in numbers
What’s real on this page: the search terms are live Google Autocomplete data collected today; the clusters, the brief, the article and the editor scores were produced with Claude; every outreach prospect is a real page that answered HTTP 200 today. The agent code that runs this on a schedule is what the four milestones deliver. Search volume and difficulty come from DataForSEO in the build, so none are shown, and the toxic-link screen uses a clearly marked synthetic sample.
The suite
Mines, filters and clusters by intent; prioritises by intent, fit, volume and difficulty; re-ranks weekly from Search Console positions.
Briefs, outlines, meta tags and articles in your brand voice — scored by an editor rubric and revised until they clear the bar.
Screens links for toxicity with evidence, proposes disavows, scouts and qualifies outreach prospects, feeds insights back.
Postgres + pgvector holds keywords, clusters, drafts, scores, links and insights — agents read and write, never call each other.
An orchestrator plans each cycle; a scheduler runs it with retries, budgets and a cost report per run.
Typed Python, documented in MkDocs — a fourth agent (technical SEO, local SEO) is one module and one schedule entry.
01 · Keyword agent
Six seed topics expanded into the long tail people actually type. The agent drops what will never convert — jobs, courses, other markets — groups the rest by intent and ranks the clusters.
Priority = 100 × intent weight × business fit × (0.85 + 0.15 × long-tail share) — × volume ÷ difficulty in production
| # | Cluster | Intent | Priority | Planned page |
|---|
02 · Content agent
The top cluster becomes a brief. A writer drafts outline, meta tags and article; an editor scores the draft against a rubric and sends it back until it clears 85. Then it lands in WordPress as a draft for approval.
03 · Backlink agent
A crawler checks every candidate, classifies it and drafts the pitch; a toxicity screen weighs anchors, link-farm patterns and topical mismatch and proposes — never auto-submits — a disavow file. What it learns flows back to the other agents.
| Prospect | Type | Why it fits | Next action |
|---|
| Linking domain | Signals | Verdict |
|---|
04 · The cycle
The agents never call each other directly — they read and write one memory, and the orchestrator decides the next step. Debuggable, resumable, easy to extend. Below are this demo’s steps, written the way each cycle’s run report will read.
Build plan
Fixed price, paid per approved milestone. You keep the code and the documentation; running costs (Claude API, DataForSEO, hosting) are billed to your own accounts.
No — it shows the outputs of one cycle and how the agents hand off, produced with Claude for this proposal. The scheduled, autonomous system with its code, tests and documentation is what milestones 1–4 deliver.
An Anthropic API key, a DataForSEO account, Search Console access for the sites, and WordPress (or CMS) credentials for publishing drafts. Everything runs on your accounts.
Only if you switch it on. By default articles land as drafts and disavow files as proposals; you can raise the autonomy per agent once you trust the scores.
Yes. Each agent is a typed Python class with its tools and a schedule entry; the documentation includes a guide for adding a fourth agent.
Claude API usage and DataForSEO requests per cycle, plus a small server. Every run report shows what that cycle cost, a per-cycle budget caps it, and it all bills to your own accounts.
You do — repository, documentation and the Loom walkthrough are handed over with the final milestone.