How Five AI Agents Publish One Article

The article you are reading right now came out of an assembly line. Not a metaphorical one. A research agent gathered the source material, a writing agent drafted the piece, a review agent checked it, an image agent produced the picture at the top, and an SEO agent audited the finished page. At our September AI workshop I walked through exactly how that pipeline works, and where a human still belongs in it.

At our August workshop, AI wrote, tested, and deployed a software feature by itself. This month I showed the other half of what we do with agents: publishing. Every article on this site, including this one, moves through the pipeline described below. I am not going to pretend a person typed it all. I am going to show you the assembly line and tell you where the person actually matters.

What does each agent in the pipeline do?

Five specialized agents each own one stage: research, writing, review, images, and SEO. Think of an agent as an AI person assigned to a specific task, with its own rules and its own procedure for doing that job.

First, a research agent goes off and gathers material on the topic. One of our engineers, Mike, runs a security agent this way, sending it out to find the latest IT security threats. Same idea here, just pointed at whatever the article covers.

When the research comes back, it gets handed to the writing agent, which builds the blog article from that material. Then a review agent goes over the draft. Are there typos? Does it need citations, and are they actually in there? Are the images in the correct spots within the article?

The image work is its own little team. One agent generates the image, and a separate reviewer agent verifies it, because AI images fail in weird ways. The reviewer’s rules include checks I demoed live at the workshop: make sure the finger count is correct, make sure there is a natural grip on objects instead of a limp hand holding a laptop, make sure glasses have two lenses and not more.

Last, an SEO agent audits the completely finished page, validating that everything on it is correct from an SEO point of view so the article can get found on Google and in Claude searches and other AI searches.

Why not just have one AI do everything?

Because handoffs and specialization are the whole point. Each agent carries a narrow set of rules for its one job, so the reviewer is not distracted by writing and the writer is not trying to remember image standards. The agents pass results to each other, and they can run in parallel. While one agent is doing research, other agents are doing other things at the same time. I can be building one blog article while research runs for the next one.

The other reason is honesty in review. The agent that made a thing should not be the one deciding the thing is good. A fresh agent with reviewer rules catches what the creator glosses over, in images and in prose alike.

How do hooks keep the pipeline from skipping steps?

Hooks guarantee that a required step happens every time a defined event occurs. When I publish an article, a hook triggers my SEO auditor automatically. I never have to remember it, and neither does the AI.

That matters because plain instructions get forgotten. You tell Claude this is something we always do, and it says, “Oh yeah, I’m sorry, I just forgot this time.” Hooks stop that class of problem. They are great for anything that must occur when something happens: cross referencing, validating, making sure a step does not get forgotten. They can also act as guardrails. A hook can stop the AI cold if it tries to delete data, forcing it to say “I can’t do that” instead. I covered that side of things in AI agents, hooks, and guardrails.

You do not hand build any of this. Tell Claude or ChatGPT what you want the skill, agent, or hook to do, and it builds the correct form and attaches it to your settings.

Should the final step auto-publish or wait for a human?

Both options exist, and the right answer depends on your confidence with your agents. After you have run quite a few articles through and read them, you may say, yeah, the AI does a pretty good job, and you develop a comfort level.

I still encourage you to manually review content before it ships. I say that from experience, not caution. I have published stuff and thought, that is not quite right, or I do not like that, or that is just flat out wrong, and I had to call it out. Claude’s response is basically “oh yeah, my bad,” which does not unpublish anything. The pipeline gets you 95 percent of the way there fast. The last look is cheap insurance, and it is the step I refuse to automate away on our own site.

What trips people up when building their own pipeline?

Two things: stale sessions and vague skill descriptions. First, anything you build, whether a skill, an agent, or a hook, will not load in the session you built it in. Close out and restart a new conversation, then verify. In Claude Code you can type the skills command to see everything currently loaded, and there are matching listings for hooks and agents. If your new piece is not listed, restart, and if it is still missing, find out why.

Second, keep every skill specific. If two skills overlap, say two different article reviewers, the description has to make the difference unmistakable: this one reviews security articles, that one reviews everything else. If the description is not specific enough, the AI will grab the wrong skill. When that happens, the fix is to go back and sharpen the description until it is unambiguous for your use case.

Build the skill, then the agent that uses it, then the hooks that validate what the agent did. That order, repeated, is the pipeline. We build the same kind of agent assembly lines for clients as part of our business automation services, for content and for plenty of processes that have nothing to do with content.

Want a content or workflow pipeline like this in your business?

The pipeline in this article is real and it produces the pages on this site. If your team has a repeatable process that eats hours every week, whether that is content, reports, or something else entirely, tell us about it. We build agent pipelines for clients the same way we built our own.

FAQs

Do the agents really run at the same time?

Yes, where the work allows it. Research for one article can run while another agent is drafting a different piece, and when the research comes back it gets handed to the writing agent. Agents pass results to each other the way coworkers hand off work, so you are not sitting there waiting on one long serial process.

What stops one agent from approving its own sloppy work?

Separation. The agent that generates an image never grades that image. A separate reviewer agent looks at it fresh, checking things like finger counts, natural grip on objects, and glasses that have exactly two lenses. The same idea applies to the article itself, where a review agent checks the writer's output for typos, citations, and image placement.

Should the pipeline auto-publish, or should a person review first?

That depends on your confidence level. After you have run enough articles through and read them, you build a comfort level with what the AI produces. Even so, I still recommend manually reviewing content before it goes out. I have published things that were not quite right, and some that were flat out wrong, and I had to call them out after the fact.

Why did my new agent or skill not show up when I built it?

New skills, agents, and hooks do not load in the session where you created them. Close out and start a new conversation, then check the skills, agents, and hooks listings to confirm they appear. If something is still missing after a restart, that is your cue to dig into why.

How is a hook different from just telling the AI to do something?

An instruction can get forgotten. The AI will apologize and say it just forgot this time, which is not much comfort. A hook fires automatically when a defined event happens, so the step cannot be skipped. My SEO audit runs as a hook every time an article publishes, and hooks can also block actions, like stopping the AI before it deletes data.