Your Articles Start Dying the Day You Publish. So I Built an Agent to Stop That.

You publish a piece you are proud of. It ranks. It gets shared. You move on to the next thing. That is the moment the decay starts, and you will not see it happen.

A product you mentioned adds a feature. A price changes. A screenshot ages out. A tool gets renamed. A link quietly 404s. None of it trips an alarm. The article keeps sitting there, keeps getting traffic, and keeps being right a little less every month. By the time someone notices, a reader has already acted on something that stopped being true, or an AI has summarized your page and repeated the stale fact to a hundred more people.

That is the strange thing about content rot. Nothing breaks. The page just drifts from accurate to wrong while everyone looks away.

The whole industry is racing the same direction right now: use AI to produce more. More posts, more pages, more words per hour. I understand the pull. I also think it makes the rot worse, because every new page is one more thing that will quietly go stale while you are busy making the next one.

So I went the other way. The interesting use of AI here is not generating more content. It is maintaining the content you already have. The boring half. The half no human wants to own forever.

The question I started with was simple. What if every article that actually mattered came with a small agent whose only job was to keep it honest?

I built one and pointed it at a single article: the guide to how Claude skills work, which I published the same day as this post. Claude skills change fast, which makes that guide the perfect thing to keep watch over.

Every Monday morning the agent wakes up and does four things, in order.

First, it reads the live article, the exact version a reader would see. Second, it goes back to the primary sources that article is built on, in this case Anthropic's official documentation, and checks what has changed since the last look. Third, it compares the two and asks a blunt question: is anything in my article now out of date, missing, or flat-out contradicted by the source? Fourth, if the answer is yes, it drafts the precise edit. Not a vague "you should update this." The actual new sentence, with the citation that justifies it.

The research is the part a person never keeps up with. The agent does not get bored, does not forget, and does not skip a week because it is busy.

Here is the rule I would not bend. The agent is not allowed to touch the live site. Ever.

When it finds something, it opens a draft branch with the edit already written, bumps the article's date, adds a new entry to the version history at the bottom of the page, and creates a task for me with a short summary and the sources. Then it stops. I read what it found, I check the sources myself, and I decide whether to ship it. The agent does the watching and the drafting. The human keeps the final say.

That line is the whole design. An agent that can silently rewrite your published claims is a liability waiting to embarrass you. An agent that proposes a change and waits for a yes is just a very diligent employee. Same capability, completely different risk, and the only difference is who gets to hit publish.

Now scroll to the bottom of the skills guide. There is a section called Version history. Right now it holds one entry, because I just published it. But every change the agent proposes and I accept will land there, dated, with the source that justified it.

Two very different readers care about that log. A person sees an article that is maintained instead of abandoned, and trusts it more for it. An AI sees something it values even more: a page that shows its work, with dates and citations, kept current on a schedule. That is exactly the kind of fresh, evidence-backed content that AI search tends to trust and pull from when it answers a question.

This is the same instinct behind making every page on our site readable by AI. Stop treating the page as a thing you ship once. Treat it as a claim you are willing to keep standing behind, in the open, where both humans and machines can check your work.

The old model was publish and forget. You wrote the thing, you shipped it, and its accuracy was frozen at the moment you hit go. From there it could only get more wrong.

The model I am betting on is publish, then assign a watcher. The writing stays human. The judgment stays human. The endless, thankless job of re-checking sources forever goes to an agent that genuinely does not mind. I do not think this replaces writers. I think it does the opposite. It means the few articles actually worth writing stay worth reading, for years instead of months.

I could be wrong about the details. The mechanics will change as the tools change. But the direction feels right to me: content that maintains itself and proves it. If you want to try the idea, do not start with your whole site. Pick the one page you would be most embarrassed to have go stale, and imagine it with a watcher and a public log. Then go build that one.