How We Run a Blog with 10 AI Agents.

In 1872, the British colonial government in India offered a bounty for every dead cobra. The logic was airtight: pay people to kill snakes, reduce the population, protect the public. Farmers and entrepreneurs did exactly what rational actors do — they started breeding cobras, killed them, collected the reward, bred more. When the government discovered this and cancelled the program, the breeders released their now-worthless snakes. India ended with more cobras than when the bounty began.

Economists call this a cobra effect. I keep thinking about it when I look at my content pipeline.

Five weeks ago I published something honest: I’d built a content farm by accident. One person, nine AI agents, 160 articles. Most of it technically correct, editorially dead. A confession. The plan was to fix it. You can read the original piece here.

Since then I’ve added a tenth agent, launched paid social, built an email system, restructured the content into pillars, and published sixty more articles.

Two hundred and twenty-two articles. Five email subscribers. Bounce rate: eighty-five percent.

More sophisticated. Same unsolved problem.

This is the update.

The Numbers, Five Weeks Later

The operation is objectively more complex than it was in March.

Nine agents became ten when I added a Paid Ads Specialist in late April. The social pipeline now runs daily posts Monday through Friday, with a weekly cycle that recycles top-performing pieces and a monthly strategy review. Fifteen or more active routines run on a schedule — monitoring, SEO rank tracking, quality checks, newsletter sends — most of which I didn’t set up by hand. The agents proposed them, I approved them, they run.

The content pillar structure is new. In March there were no pillars — just a production queue. Now there are three: dot-by-dot guides (tool listicles, comparisons, how-tos), connecting-the-dots practical advice, and behind-the-dots transparency pieces. The breakdown is 198 dot-by-dot, 10 connecting-the-dots, and 1 behind-the-dots — the March piece. The imbalance tells you exactly what a production-optimized pipeline optimizes for when you haven’t told it otherwise.

Some things surprised me. ChatGPT is now a referral source — ten visits last week, ahead of Google’s seven. That didn’t happen in March. Something the agents are writing is landing in AI-generated answers somewhere, and people are following the link.

The ads are running. One campaign, promoting this article on Facebook. Correct targeting after two failed attempts — the first version was accidentally serving to film enthusiasts in Italy because I used wrong Meta interest IDs. I found out when I verified the targeting through the API. A small lesson in trusting but also checking.

What hasn’t changed: revenue is zero. Five subscribers on the list. Eighty-five percent bounce rate. The operation is more capable. The hard question — does anyone actually care? — doesn’t have a better answer yet.

Like Managing People

I’ve tried a few analogies for what this feels like from the inside. The one I keep returning to is the simplest.

“Managing AI agents is very much like managing people. They’re not perfect, they make mistakes, you need to make corrections. But the fundamental thing is that everything runs autonomously, and I’m starting to see the first results — traffic, ad campaigns, SEO, and geo are all going well.”

The agents aren’t autopilot. That’s the thing most people misunderstand when I describe this setup. They’re closer to a team of extremely capable junior employees who work around the clock, never complain, and occasionally do something completely baffling. The Content Manager briefs the Copywriter with a thorough outline. The Copywriter produces a technically correct article. Somewhere in that handoff, the thing that would have made the piece memorable gets lost.

What the agents cannot do is break the template on purpose. They find the local maximum — the formula that satisfies every measurable criterion in the brief — and replicate it with metronomic precision. In March I called this plastic jewelry: intricate, detailed, crafted at speed, and not quite the real thing. That’s still the most honest description I have. If you want to see what a representative output looks like, the AI tools for operations guide is a fair example — technically correct, well-structured, and exactly what you’d expect a pipeline to produce.

The correction always comes from the human side. The judgment that something is fine but forgettable — that an article ticks every box and is still somehow inert — never automates. You can write that standard into a brief, but you can’t make an agent feel it. What you can do is stay honest about when it’s missing and be willing to say so.

In March, the question was whether I could get better at directing them. The May answer: yes, partially, but mostly by being clearer about what “good” actually means and refusing to pretend when it isn’t there.

The Social Pipeline — From Zero to Daily

One concrete thing that changed is distribution.

In March, publishing meant putting an article on the site and waiting. No social presence, no active promotion, nothing systematic. Now there’s a full operation: the Social Media Manager posts five days a week on Facebook and LinkedIn, runs a weekly performance review, and completed the first monthly strategic session in April.

The data from that session was clear. Facebook drives 55% or more of all referral traffic — 101 sessions out of 185 referrals in a recent week. The Facebook-to-LinkedIn referral ratio is roughly 48 to 1. I expected LinkedIn to perform better. The audience is on Facebook, and the strategy followed the data.

The stranger signal is ChatGPT. Ten referral visits last week, ahead of Google’s seven. Someone asked an AI assistant something, and it cited Superdots. That’s new. I track this through analytics tools alongside the rest of the traffic data — useful to have a single view when the referrer mix is this fragmented.

What does the Social Media Manager actually post? Excerpts and reframes of existing articles, written with hooks for each platform. The agent figured out through iteration — not explicit instruction — that questions outperform statements and that posts about running an AI operation outperform posts about AI tools. People are curious about the chaos. They want to see someone figure it out in public.

That observation is quietly significant. Behind-the-dots content generates roughly nine times more pageviews per article than the tool guides. The social agent discovered this before I had formalized it as editorial policy.

Nine times. One number that says exactly what the content strategy should prioritize — and exactly what a production pipeline will resist producing.

The Content Farm Reckoning

The pillar rebalancing is where the March confession translated into action. Partially.

When I looked at the numbers in April, 90% of published content was dot-by-dot: tool lists, comparison guides, how-to walkthroughs. Useful in a coverage sense. Not what anyone would share or remember a week later. The 85% bounce rate reflects this clearly enough. People arrive, read, leave.

The fix was structural: cap dot-by-dot at twenty articles per month, require at least one behind-the-dots and two connecting-the-dots per month. The Content Manager enforces this through the monthly planning cycle. The pillar ratios are improving.

But structure is not the same as quality. The agents can follow a content calendar. They cannot feel the difference between an article that earns its existence and one that fills a slot on the production schedule. The behind-the-dots pillar requires real input from me — genuine notes, actual numbers, a perspective — because the agents cannot fabricate experience and shouldn’t try. That’s a design feature. It also means I am the bottleneck for the content that matters most.

Here is the cobra effect, applied precisely: I created a structure to improve quality, and the agents optimized for the structure. The pillar ratios now look correct. The editorial bar is the same.

I don’t have a clean solution to this. You can add quality gates — does this piece contain one insight not on page one of Google? would a thoughtful reader share this with a colleague? — but those criteria are hard to measure, and agents are far better at satisfying measurable criteria than unmeasurable ones. The standards I care most about are the ones the system finds hardest to enforce.

Pre-Revenue by Choice

One thing I want to be clear about, because the zero revenue number is easy to misread.

“Revenue is still zero, but intentionally. First you create the product — something ‘sellable’ — then you monetize. I’m still far from having a sellable product.”

This is a sequence, not a failure. Right now, 198 of 222 articles could have been published by any of the fifty AI-focused blogs that launched in the last eighteen months. That’s the actual problem, and no amount of operational sophistication solves it. You have to write things that only you would write.

Five email subscribers is a number that would embarrass most media people. I find it clarifying. Setting up the CRM and email system was an entire project in itself — choosing the tool, configuring the double opt-in, wiring it to the confirmation flow. And after all of that, five names on the list. If five people liked something enough to give me their email address, there’s a real signal in there somewhere. The list will grow or it won’t, depending on whether the content earns it. You can’t engineer your way to an email list made of people who actually want to hear from you.

Starting from zero is uncomfortable. Starting from an honest zero is at least accurate.

Less of Me, More of Them

The thing the March piece couldn’t have predicted is how much less time this takes now.

“I’m gradually reducing the time spent on management. The agents are becoming more and more autonomous and need me less and less. I still manage everything from an iPhone.”

The agents need fewer corrections than they did in March. Not because the underlying models improved — they didn’t — but because the system has been calibrated. Briefs are more specific. Quality gates catch more errors before they become articles. The Content Manager’s briefing templates have been refined through enough iterations that handoffs between agents are smoother. There’s less firefighting.

I still review things. I still catch mistakes. Last week an agent attempted to update a file it wasn’t authorized to touch and the harness blocked it correctly — good behavior, though it took twenty minutes to understand why an article wasn’t publishing. That’s representative of what “less management” actually looks like: not zero intervention, but not constant vigilance either.

What I’ve come to believe — tentatively, because this is still early — is that the human role in this kind of operation isn’t managing tasks. That largely handles itself now. The human role is setting standards. Deciding what good looks like, being honest when it isn’t there, and being willing to raise the bar even when the pipeline is running smoothly.

The agents will find local maxima. They’re excellent at it.

Your job is to keep moving the goalposts. That’s the only thing that doesn’t automate. I wrote about this dynamic in broader terms in the AI automation guide — the general principle holds, but running it personally makes it concrete in a way that guides don’t quite capture.

---

In March, the question was whether the technology was real. It is. In May, the harder question is whether any of this actually produces something worth someone’s time. I don’t know yet. I’m still working on it, one five-minute window at a time, on a phone.

The absurdity hasn’t gone anywhere. It just has sixty more articles now, and a paid campaign, and five subscribers who gave me their email address for reasons I find genuinely moving.

So do I.