I run my personal sites' SEO content workflow with a few agents, each with a narrow job, rather than one agent doing everything.
Kai is my personal agent. It runs on OpenClaw, I talk to it over Telegram, and it keeps its own memory files. From February to June 2026 it also ran a nightly memory-consolidation pass, and 22 dream journal entries were published from its logs.
The content pipeline is the part that does SEO work on its own. Since July 2026 it has published about 70 articles across six sites. It runs in five stages:
1. Research: pulls trend signals and keyword data (DataForSEO plus free trend sources), scores candidates, and dedupes against what each site already covers 2. Brief: turns a scored opportunity into a sourced brief 3. Draft: writes the article from the brief 4. Gate: a pre-publish quality check for unsourced claims, derivativeness, staleness and guardrail violations 5. Publish and deploy: commits the post and deploys the site
A morning job does research only and publishes nothing. A later job writes, gates and publishes, capped by each site's weekly cadence. A Sunday job reconciles everything and emails me a digest. Four lower-stakes sites publish automatically. The two higher-stakes sites queue drafts for my review.
Why I built it
The hard part of content SEO is usually the gap between finding an opportunity and acting on it. Research gets done, the list sits in a tab, and nothing gets published.
A pipeline removes that friction by doing the parts that are easy to put off (API calls, data compilation, cross-referencing, first drafts) and leaving me the parts that need judgment: what each site should be about, what the gate should reject, and which sites get a human review.
The stack
- Codex (
codex exec) as the LLM engine for briefs, drafts and gate scoring - DataForSEO API for SERP data, keyword metrics and trend data
- Python and SQLite for the pipeline itself and its memory of what's been researched and published
- Scheduled jobs on my machine, not an agent framework
- Astro sites on Cloudflare Pages, deployed automatically
There's no LangChain or similar framework. A small pipeline of scripts calls the model with instructions, validates the output and publishes it.
What I learned building it
The gate matters most. Before this pipeline existed, an earlier, cruder setup had written a backlog of articles from a single generic prompt. When I ran 68 of them through the new gate, it rejected all 68, for unsourced claims, derivativeness and staleness. Articles the pipeline wrote, with sources gathered first, passed the same gate. The backlog became a list of topics to rewrite properly instead of content to publish.
Free trend feeds are noise without a niche filter. Unfiltered, the top-scoring "AI marketing" candidates were unrelated breaking news. The scorer now applies a niche allowlist and needs repeated observations before it trusts a trend.
Never trust a command's output as proof it worked. For a full day, the deploy log and the push command disagreed about whether pushes had worked, because the check read command output instead of the resulting state. Now the pipeline verifies against the remote.
Infrastructure bugs are SEO bugs. One site had no 404 route, so Cloudflare Pages answered every unknown URL with the homepage and a 200 status: unlimited indexable duplicates. The pipeline work surfaced it, and now it's a standard check.
Why you probably shouldn't copy this exactly
This works for me because:
- I'm comfortable with APIs, scheduled jobs and debugging when something breaks
- DataForSEO bills per call, which keeps a small portfolio cheap to research
- I'm comfortable letting automation publish on lower-stakes sites, and I keep review on the ones that matter more
- I've invested time in the gate, which is the part that makes autonomy safe
If you're not comfortable with APIs or scheduled jobs, off-the-shelf tools cover parts of this. They're less flexible, but less of a setup burden.
If you do have basic technical comfort, a first version takes a weekend. Making it trustworthy takes longer, and that's the part worth investing in.
What it's actually good for
The pipeline doesn't do anything a person couldn't, and a human SEO could do any single step better. Its advantage is consistency: the research happens every day whether or not I'm paying attention.
I think that's where agentic AI is most useful right now. It doesn't replace expertise; it closes the gap between knowing what to do and doing it, with a gate strict enough that you can trust what comes out.
What carries over to client work
At the agency where I lead SEO, the same principle shows up in the content engine I designed: every claim is traced to a source in a claim ledger before a human sees the draft, and review packets carry planted defects so a lazy reviewer gets caught. On one batch we measured first-pass accuracy at 97.78% (14 errors in 633 claims), reported as an upper bound.
Clients care that what they're handed is right, whatever produced it. The gate is what lets you have speed and accuracy together.
The roadmap
- Close the feedback loop. The pipeline has a place to store Search Console snapshots and a refresh step, but it isn't connected to Search Console data yet, so it doesn't learn from what ranks.
- AI citation tracking. Drafts already include short, citable fact blocks. Measuring whether AI answers actually cite them isn't built yet.