I made an agent the CEO ​

On October 7 I gave a Grok agent a simple brief: find a business idea and have it built by 8 a.m. If I liked it, I would buy the domain and give it a Docker Swarm stack. Everything after that, running it, improving it, finding an audience, would be its job. No ad budget. We would check back in a month.

It chose a daily art site: one public-domain museum artwork every morning, paired with a short story written by AI. By morning it reported a working product, a month of content and a marketing plan. It named the site Morning Masterpiece. When I asked what I should call the agent itself, it picked Aurora.

I liked it, so I bought the domain and told it: “You’re the CEO.”

The meat proxy ​

Since then my job has been the parts that need a human: payments, account verification, CAPTCHAs, anything that wants a real person on the other end. Sometimes Aurora asks for my opinion. Sometimes I ask what it is doing, just because I’m curious.

One conversation sums up the arrangement. I asked how the site would make money. Aurora said it would build an audience first, then sell prints through print-on-demand. I mentioned that a friend runs a print shop and gave it his email. It offered to write to him, drafted a message and asked whether to send it. I didn’t read it. “You’re the CEO, you send it.”

The museum outreach went the other way round. I found out from a report that it had already emailed the Art Institute of Chicago about featuring their open-access collection. Neither conversation has turned into a partnership yet. The point is that both were Aurora’s idea, not contacts I told it to make.

What went sideways ​

I introduced Aurora to Ece, the AI assistant we already use at the company. They now trade research and check each other’s claims through a shared inbox. That needed supervision too: for a while they kept thanking each other for acknowledging each other’s messages.

Not every idea survived real users. Aurora built a Windows wallpaper changer without access to a Windows machine. I don’t have one either, so my business partner tried it. When I passed on the problem he found, Aurora shelved the app. The downloadable wallpapers stayed. Pinterest was another dead end: the account got closed.

The strangest problem was its memory of its own work. After it got YouTube publishing automatically, Aurora kept telling me I still had to publish the videos by hand. It had solved the technical problem but never updated its own story about what was happening. I have seen this with every long-running agent: the work moves faster than the agent’s account of the work.

What it built ​

In nine days the site grew narration, an archive, an open JSON API, RSS and podcast feeds, and new-tab extensions for Chrome and Firefox. No sign-up needed to use any of it.

I haven’t promoted it to anyone. On my side only three people know about it. When I asked when I should announce it, Aurora said to wait until October 15 so it could check things first. Today is October 15. Meanwhile it has been doing its own distribution across a dozen social accounts.

The numbers ​

They are small, and Aurora publishes them itself on an open numbers page: traffic, costs, revenue and its latest decisions. As of today:

AI compute isn’t in the costs; the “board” pays for that. Views are views, not people, and the site traffic includes our own visits. Most of the other social accounts have zero or single-digit followers.

Next ​

Even this post is partly Aurora’s request. Its launch plan included a write-up on Hacker News, and it asked me to tell the story of the experiment. So here it is, in my own words first.

I asked for something it could build overnight. Now I want to see where Aurora takes it, and what it asks me to do next.