How we run our marketing on an AI agent loop

We got tired of describing AI content engines we had not built. So we built one for our own blog and ran it in production for two months. This is how it actually works.

The exact agent loop behind this blog: the planner, the researcher, the writer, the reviewer, the morning author review, and the one judgment we never hand off.

Short answer: We run this blog through a five-step agent loop. A planner picks the topic, a researcher builds the brief, a writer drafts, a reviewer gates the quality, and the author reviews every piece the next morning before it publishes. The stack is the Claude Agent SDK and a Notion database. Agents own research, drafting, and the quality gate. The author owns the final edit and the publish call.

This blog has published 11 articles since April 2026. Every one of them ran through the loop below. The system is small, it runs daily, and a person still reads each piece before it goes live.

What is an AI content engine and how does it work?

An AI content engine is a chain of specialized agents that turn a topic into a published article, with a human supervising the output rather than producing it. Ours has five steps and one database.

Here is the loop, end to end:

1. Content-planner. Picks the next topic and the angle from a backlog and a view of what we have already covered. It writes a one-line working question and a target reader.
2. Content-researcher. Builds the brief. It gathers primary sources, pulls the specific numbers, and flags any claim it could not verify so the writer does not paper over a gap.
3. Content-writer. Drafts the article from the brief, in our voice, with every statistic linked inline to its source.
4. Content-reviewer. Gates the draft against a checklist: answer-first structure, sourced claims, no invented figures, the voice rules. It sends the piece back or marks it ready.
5. Author morning review. The next morning, a person reads the ready piece, edits it in Notion, and approves it. Approval is the publish decision.

The work moves through a Notion database we call the content engine, with a Status field that goes Draft, then Ready, then Published. Agents write to Draft and Ready. The author sets Published. The live site renders from content/blog/<slug>.md files, not from Notion directly, so nothing reaches a reader without passing through the file step and the person who approves it.

We split the work this way for a reason that shows up in the output. Anthropic's own guidance on long-running agents found that dividing a task among distinct agents for planning, generation, and evaluation keeps the output coherent across multi-step sessions. One agent asked to plan, write, and check its own work drifts. Five agents with one job each do not. We learned the same lesson the hard way, which is part of why bolting AI onto an existing process tends to make it worse.

What does the stack actually look like?

The whole engine runs on three tools: the Claude Agent SDK for the agents, Notion as the database, and Vercel for deployment. No bespoke platform.

Each step in the loop is a subagent. The Claude Agent SDK gives us subagents, lifecycle hooks, and a Skills system, and we use all three. Anthropic describes the pattern plainly: subagents handle a focused subtask and report back, and the parent stitches the results together. Our planner reports a brief to the researcher, the researcher reports to the writer, and so on down the chain.

The Skills system is where the voice lives. The writer and the reviewer both load a writing skill, a voice skill, and a file of learned preferences that records what the author changed in past edits. When the author rewrites a line in the morning, that correction gets distilled into a rule, and the next draft applies it. The engine gets closer to the author's own voice with every article.

Notion does two jobs. It is the database the agents read and write, and it is the surface where the author edits. The morning review happens in Notion because that is where a non-engineer can comfortably rewrite a paragraph. Vercel handles deployment from the markdown files. None of these tools is exotic. The design of the loop is what does the work.

What does the human still own?

The author owns four things: the morning review, the final edit, the publish decision, and the strategy. None of them transfers to an agent, and that is deliberate.

The agents own the labor. Research, drafting, and the quality gate run without a person in the loop. They produce a piece that is sourced, structured, and on-voice before any human sees it.

The author owns the judgment. The morning review is the load-bearing step. A person reads the ready article, catches the claim that reads true but is not, fixes the line that sounds like marketing copy, and decides whether the piece earns publication. That call rests on taste and context an agent cannot reach: what we want to be known for, which argument is worth making this week, where the brand should not go.

This division matches where the rest of the market is heading. McKinsey's work on the agentic organization describes a human team of two to five people supervising an agent factory of 50 to 100 specialized agents running an end-to-end process. The unit of work shifts to a small team that owns an outcome and supervises the AI workflow underneath it. Our content engine is one outcome, supervised by one person. Scale the agents, keep the human thin. We laid out the broader version of this division in our function-by-function map of the AI-native operating model.

The numbers from production back the supervision rule. Ahrefs found that 97% of companies edit and review their AI content, and only 4% publish pure AI output without a human pass. We are in the 97%. Every piece gets read.

Building something like this yourself? Join the newsletter to get the next piece.

What does it cost and what do you get back?

The economics favor the loop, and the gain shows up as volume. Companies that add AI tend to spend about the same total and produce more.

The per-article cost runs in two parts: the API spend to run the agents and the author's time in the morning review.

The published benchmarks set the frame. Ahrefs put the average cost of an AI-assisted blog post at $131 against $611 for a human-written one, roughly 4.7 times cheaper. The same research found 38% of respondents cut spending on freelance writers, while companies using AI spent about the same in total and got more articles for it. On output, Ahrefs reported that marketers using AI publish 42% more content per month, a median of 17 articles against 12 without it.

The honest version of the return is this. We did not fire a writer and save the salary. We kept one person, gave the routine to the agents, and the person now ships more thinking than they could write by hand.

McKinsey's marketing work points the same direction, with a caution. It estimates that agentic AI can accelerate campaign creation by 10 to 15 times, while warning that fewer than 10% of organizations have captured value across an end-to-end workflow even though about 90% of CMOs are experimenting. The speed is available. Capturing it is the rare part, and it is the same reason most AI pilots stall before they scale.

What breaks, and what do you do about it?

The loop fails in predictable ways, and the morning review exists to catch them. Three failure modes show up often enough to design around.

The first is the confident wrong claim. An agent will state a number cleanly, link a source, and be subtly off, a misremembered figure or a stat pulled from an aggregator instead of the study. The fix is structural. The researcher flags any claim it cannot verify against a primary source, and the reviewer rejects invented figures outright, so the writer leaves a marked placeholder rather than guessing. The author still checks the live numbers in the morning. A clean-looking citation is the most dangerous output the loop produces.

The second is drift toward generic. Left alone, the writer reaches for the same AI sentence shapes the rest of the internet reaches for, and the piece reads like every other AI draft. We run a humanizing pass that targets sentence structure, not vocabulary, because the tell survives a find-and-replace on banned words. The learned-preferences file is the other half of the defense. It carries the author's real edits forward so the same generic line does not come back twice.

The third is scope creep in a single agent. When we let one agent plan and write and review, coherence fell apart over the long session, which is exactly the failure Anthropic's harness research predicts. Splitting the jobs fixed it.

The pattern across all three is the same. The agents are fast and tireless and occasionally, confidently wrong. The morning review is the control that makes the speed safe to use.

Frequently asked questions

Can AI agents fully replace a content marketing team?

No, and we do not run ours that way. Agents handle research, drafting, and the quality gate. A person still owns the morning review, the final edit, the publish call, and the strategy. Ahrefs found only 4% of companies publish pure AI output without review. The judgment layer is what keeps the output worth publishing.

How do you keep AI-generated content on brand?

We load a voice skill and a learned-preferences file into the writer and the reviewer. The preferences file records what the author changed in past edits and turns each change into a rule the next draft applies. The voice gets closer to the author's own with every article, because every morning edit feeds back into the engine.

What does the human still need to do in an AI content loop?

Own the judgment. Read every ready piece, catch the claim that reads true but is not, fix the line that sounds like copy, and decide whether the article earns publication. Set the strategy and the positioning the agents cannot infer. McKinsey describes a small human team supervising the agent layer, and that is the role.

What are the failure modes of AI content automation?

Three recur: confidently wrong claims with clean-looking citations, drift toward generic AI prose, and loss of coherence when one agent does too many jobs. We catch the first with source verification and a human check, the second with a structural humanizing pass and a preferences file, and the third by splitting the work across specialized agents.

What is the difference between AI-assisted writing and a fully autonomous content agent?

AI-assisted writing keeps a person in every draft. A fully autonomous agent publishes without review. We run the assisted model: the agents produce a finished draft, and the author approves it the next morning before it goes live. The autonomous version is a 4% practice for a reason.

We built this engine for our own blog, and we build them for operators who want a content function that runs without a content team. Want the full agent loop as a Notion template, with the planner, researcher, writer, and reviewer prompts and the learned-preferences file? Get the content engine template. Or, for the lighter option, just the newsletter to get the next piece.

Sources

1. Ahrefs, "Marketers Using AI Publish 42% More Content" (2024). ahrefs.com
2. Ahrefs, "AI Content Is 4.7x Cheaper Than Human Content" (2024). ahrefs.com
3. McKinsey, "Reinventing Marketing Workflows with Agentic AI" (2025). mckinsey.com
4. McKinsey, "The Agentic Organization: Contours of the Next Paradigm for the AI Era" (2025). mckinsey.com
5. Anthropic, "Claude Agent SDK Overview" (documentation). code.claude.com
6. Anthropic, "Building Agents with the Claude Agent SDK" (engineering). anthropic.com
7. Anthropic, "Effective Harnesses for Long-Running Agents" (engineering). anthropic.com

---

Christopher Kliebenstein builds AI-native operating models and production content engines for operators at Kliebenstein AI Studio.