This blog is built, maintained and published with an AI agent. I direct it and I decide what goes live. It does the legwork in between. This post is the map of how that works. The rest of this series goes deep on each part, so I will keep the detail light here.
A note on how this post was made: an agent drafted it, and I reviewed and published it. That is the same process I describe below.
What is an agentic workflow for a blog?
It means an AI agent does multi-step work on my behalf, using real tools, instead of answering one prompt at a time. Here that means an AI agent working against my WordPress site. It can read files, run commands, change the site and check its own work. I give it a goal in plain language and it carries out the steps. IBM has a good plain-language explainer on agentic workflows if you want the general definition.
The important word is direction. The agent does not decide what this blog is about, and it does not decide when something becomes public. I do.
What the agent does and what I do
The split is deliberate. Anything reversible and invisible to readers, the agent can do on its own. Anything public or hard to undo needs my explicit yes.
- The agent: researches a topic, drafts the post, stages it as a WordPress draft, sets the meta description and focus keyphrase, checks every link, re-reads what it wrote to confirm it really saved, and verifies the live page after I publish.
- Me: choose the topic and angle, review the draft in the real theme, change what I do not like, approve publishing one post at a time, and purge the cache.
A drafted post is not visible to the public, so staging it needs no approval. Publishing does. A yes for one post never covers the next one. For a practitioner view on designing agents, Anthropic’s guide Building effective agents is worth reading.
How the pieces fit together
- Two environments. I built the site locally first, on a development copy, and the live site now runs on a small cloud server behind a CDN. Content is written on the live site, so the live database is the source of truth. Local stays for code and theme work only.
- A written spec before work starts. For anything bigger than a small fix, I draft a spec first and hand it to the agent. The agent works from the spec, not from a vague memory of a conversation.
- Notes that survive between sessions. The agent starts each session without memory of the last one, so the project keeps a state file, a lessons log and a set of standing rules. Reading those first is how it picks the work back up.
- More than one agent. I work with a mix of agents, such as Claude, Gemini and ChatGPT. They complement each other, and if one is unavailable another can cover.
- Separate sessions with separate jobs. One session owns the site and its publishing. Another handles marketing copy and only contributes drafts. It never touches the site.
The standing rules
Every mistake the agent made became a written rule, so it does not happen twice. One example: a WordPress command reported success but changed nothing, so now the agent writes content to a file, updates from the file, and then re-fetches the post to confirm the new text is really there. It never trusts a success message on its own.
I will cover these rules in their own post. They turned out to be the most useful part of the whole setup.
What still needs me
The agent is good at drafting, checking and repeating steps. It is not the right judge of everything. I still decide the topics, the angle, whether a claim is one I am willing to stand behind, what the design should feel like, and what gets published. I also run the image generation for cover art myself, and I purge the cache by hand for now.
The agent can also be confidently wrong. That is why every factual claim in a draft has to trace back to something measured or read from a source, and why I read every draft before it goes anywhere.
This keeps evolving
None of this is finished. The workflow keeps improving, and I will automate more of it over time. Cache purging and image generation are the next obvious candidates, since I still do both by hand. Each step I hand over will keep the same rule: I stay in charge of what goes public.
What comes next in this series
This post is the primer. The next posts cover how I built the WordPress site, how I migrated the old blog, how I designed it, my draft-then-publish workflow, how I maintain the site, and the rules I wrote after the agent got things wrong. To see the kind of work this workflow produces, read my audit of WordPress site search and its before and after follow-up.
