My last email to you went out on March 27. It is now late October.

You deserve an explanation. You also deserve more than an apology, because a lot happened in those seven months and most of it is exactly what this newsletter is about: one person running several businesses with a team of AI agents, on a budget that would embarrass a real company.

So this issue has two parts. First, why you heard nothing. Then, what the agents and I built while you were not hearing from me.

Why it went quiet

The silence came in three phases, and only the first one was a choice.

Phase one, late March to mid May: I turned it off. The fleet had started publishing too often, some days four times, and burning through quota. I shut it down on purpose. That one is on me, and I would do it again.

Phase two, mid May to early September: the gate ate everything. I switched the writing agent back on in May. It wrote every day. By September it had produced 141 drafts. Not one reached you.

Every draft passes through a safety scanner before it can be published. The scanner looks for things I cannot afford to leak: a customer identified by implication, a private number dressed up as a real figure, a quote nobody said. If the scanner cannot run, it blocks the draft. It fails closed, and I still think that is right.

The model behind that scanner sat behind one paid API account. That account ran dry in March. From then on, the scanner blocked every draft, correctly, for three and a half months. The gate worked perfectly. It just had nothing to think with.

The worst part is that every alarm I built fired. The agent logged its own blocker almost every day, accurately, to a dashboard I built for exactly that. The schedule stayed green. Drafts appeared on disk on time. If you had asked me in July whether content was running, I would have said yes and shown you proof.

A working agent producing unusable output looks exactly like a working agent.

Phase three, September to now: the bottleneck was me. I fixed the scanner in September. I also set a rule: the agent sends me one article to approve at a time, and nothing else goes out until I decide. That rule worked. The agent sent one draft and waited. I did not open it for six weeks. The system did what I asked, and the person in the loop was the slow part.

That one is also on me.

What got built while I was quiet

The writing stopped reaching you. The building did not. Since March 27 the main codebase took 787 commits. The fleet is now a bit over two dozen agents, and about a third of them are deliberately asleep.

These are the five things worth telling you about.

1. An agent that plans my day while I sleep

This took more work than anything else this year, about 190 commits on its own.

Every night at 11pm an agent reads everything that is waiting on me. That includes open work sessions, blockers the other agents raised, and decisions in the queue. Then it writes my next day into Google Calendar as real time blocks. I wake up to a plan instead of a pile.

Since April it has run 195 times and placed 2,860 blocks. It also taught me the same lesson the newsletter did. One night in October the machine rebooted in the afternoon, a tool the planner reads from was down, and the planner logged zero errors and planned nothing. I woke up to an empty calendar and a clean log.

"Zero errors" and "it worked" are different claims. The fix was a morning catch-up job that checks whether today is empty and replans it. It is a small piece of code, and it would not exist without that empty morning.

2. Agents that use my accounts without logging in

I ran an audit and found something embarrassing. Over 90 days, my agents had opened the Zoho Books dashboard in a browser about 5,700 times, mostly just to switch between company accounts.

So I stopped letting them use the browser for that. Each service now has a small tool with one API key per account and an account picker. The agent says which company it means, and the tool does the rest. Zoho, PayMongo, Cloudflare, Supabase and Apify are on it, plus a test-user tool for EsthetiqOS.

The tool that touches money got the most care. Live payment writes are off by default. When I turn them on for a business, they are capped per day and limited to a short list of actions. Agents should never type passwords, and they should never be one bug away from moving cash.

3. A software factory for EsthetiqOS

EsthetiqOS is my clinic software for aesthetic and dental practices. It has a steady stream of small tickets, the kind that eat a founder's evenings.

So I built a factory. I hand it a ticket. It makes an isolated copy of the code, builds the fix, proves it, runs a review loop, and opens a pull request. Each one carries a card that says what was proven and what was not. I make one decision: merge or not.

The first ten runs all merged, with zero reverts, and the last seven needed nothing from me until the merge click. The bottleneck moved somewhere I did not expect. The factory is not the slow part anymore. The slow part is finding tickets that are safe to hand it. Anything touching the database schema or security stays human.

4. Tarong, and agents that find the work

Tarong (tarong.dev) is new. It is an AI dev shop, and it is the most direct test of whether this whole setup can earn money and not just save it.

The agents found local businesses with no real website, mostly vet clinics, dive shops and care homes. Then they built each one a concept site to show what theirs could look like. There are 37 of them live now. Generating the images for the first 30 cost about five dollars.

Here is the part I want you to hear. After the builder agents finished, I ran a separate set of fact reviewers over the sites. They found about 60 overclaims. One good review turned into "clients rave," or a business got opening hours it never published. The builders had checked their own work and missed every one. Fresh eyes are not optional.

There is also a scanner that reads Upwork job posts, filters them, scores them and drafts proposals. In its first week it read 353 posts, kept 62, and drafted 5. That ratio feels right to me. I want it picky.

5. AIChatExport gets a nightly digest

AIChatExport is a small tool for exporting AI chat history. Its users hang out in forums and subreddits where people keep asking how to export a conversation.

An agent now finds those threads every night and sends me a short digest on my phone. Each item has the link and a reply I can paste. I post the replies myself. The agent finds and drafts. A human shows up in the thread. That split is deliberate.

The part that fits the name of this newsletter

This newsletter is called The $200/Month CEO. For most of this year that name was only half true.

The agents themselves ran on my Claude subscription. But the cheaper background jobs, like classifiers, scanners and summaries, ran through metered API accounts. Those accounts are exactly what broke. The dry account that blocked 141 drafts was one of them. This month a second metered account ran dry and knocked scheduled jobs over again.

This week the fleet moved its main routes onto the subscription first. Metered providers are now only the fallback. When a prepaid balance runs out, the system gets a little slower instead of stopping.

Flat cost is not free. The real limit now is usage quota, and the fleet shares it with my own work. So every job tags itself and lands in a usage ledger. Instead of watching a balance drain, I can see which job is spending my quota.

What happens now

The scanner works. The approval rule works. The backlog is close to 150 drafts, and some of them are good. Each of the five things above gets its own deep-dive issue, so that is where the next few emails come from.

I am also fixing the part that failed last. The agent will still send me one article at a time. But a draft that sits unanswered should not be able to stall everything quietly for six weeks. A silent queue is another empty table, and I have had enough of those.

If you subscribed in the last seven months and wondered whether this was dead: for seven weeks it was switched off. For the rest, it was writing every day into a folder, waiting on a gate, and then waiting on me.

Back now.