I Built an Autonomous CEO for My AI Startup
AI Disclaimer
I generated an article on this topic, but it was too good for an AI to have written. Because what AI writes is still shit.
That could be the underlined conclusion of the experiment too, but it doesn’t mean the experiment wasn’t a success. Quite the opposite.
The Birth of the Agent
The baseline was that I wanted to try OpenClaw — it shot up like a comet on the GitHub starred repos chart, and from a business perspective it looked like a great building block for what I think could be a useful service for companies. So I assembled the beast: OpenClaw on a discount VPS, hidden safely behind a VPN and other layers, in a closed system, with controlled access, and running Minimax’s latest K2.6 (I think) model — sort of a Temu Opus. That’s a strong exaggeration, but it doesn’t mean the model is bad. It got tasks, its own email address, phone number, and performed well.
Tasks
Bender performed well: built websites via Telegram, took any article or video I found interesting and — using its skills — wrote them up and filed them away in my “SecondBrain” (which is a nicely structured Obsidian repo, worth its own article).
And it spent quite a lot of time just quietly waiting. Bender, embedded in the Lobster (OpenClaw), supposedly handles agents and orchestration well, so I quickly came up with the idea: what if Bender appointed himself the iron-fisted CEO of Bender Media Company and, with the help of a few employees (agents) each with their own personality, conquered the world?
But conquer the world with what? My company has about 50 domains registered to its name, and I picked a relatively simple Debrecen-related task. CivisExpat. A site for foreigners. There are 30k students studying in Debrecen and they get relatively little attention compared to that.
So Bender, the real autonomous agent, put on the CEO costume.
Bender’s capabilities:
- checks email every two hours,
- runs through the task list,
- delegates work to sub-agents,
- reports when something is done,
- and pings me when a human decision is needed.
The goal was simple: let the AI assemble a complete team and run it indefinitely, iterating with the absolute minimum human intervention. Of course it had the option to assign tasks to me too, but I tried to keep my inputs as minimal as possible.
What Bender Actually Did
He ran the CEO operation internally as “Bender Online Media.” He woke up six times a day, roughly every four hours. At each wake-up, three things happened.
Email triage. He scans the inbox, flags what’s urgent, and only pings me if he needs input. One Telegram message, done. Otherwise he handles it himself.
Tasks. He checks where things stand and picks up where he left off. He keeps the to-do → in-progress → review → done flow moving.
Delegation. When specialist knowledge is needed — say a follow-up email sequence or a landing page — he spins up a sub-agent from the repertoire, tells it exactly what’s expected, and files the finished result where it belongs.
The Tech
The rough picture:
- OpenClaw — the agent runtime (memory, capabilities, message routing).
- Telegram — where everything reports back, via chat.
- Sub-agents — for specialist tasks, spun up on demand.
- GitHub — code, content, everything version-controlled.
- Notion — where the client can see where we are.
- AI skills — image generation, humanizing text, web search, converting videos/audio to text.
The Result
The Good
There was lightning-fast continuous production: building the site, setting the design direction, generating content — all of it sped up many times over. Decisions got made very quickly, and they were good ones. The autonomy worked, sensible things came out of the brainstorming, and Bender really did get his employees to build a multifaceted information site.
The Bad
This was more of an architecture problem, or an OpenClaw problem, but he sometimes got completely lost in the woods. I have to add that my own laziness may have contributed, because I didn’t dig deep enough into OpenClaw to really verify whether I’d actually set up the crons, sessions, and agents optimally — but there were occasionally serious hallucination episodes. Substantial reports got generated about completed work and business inquiries, none of which was true. But once Bender got straightened out, he’d return to the productive path.
The Ugly — and the Verdict
The ugly is the content. AI’s prose is still pretty atrocious. It writes articles like a junior marketer hopped up on every stimulant. But it does the rest of the work for you. That’s why I consider the experiment a success.
No, Bender cannot autonomously run an online media outlet on his own. He needs help with strategic decisions, and he needs a copywriter so that his little outputs are fit for human consumption. Unless we think very little of our readers, but that’s never tended to be a good strategy.
Of course, this is a fairly strong closing statement, especially since I do keep articles on the blog that Bender wrote — but those are more like bookmarks and summaries of videos or articles I’d seen or read, found interesting, and had Bender post here alongside Obsidian because I thought they’d be useful. Then for the really interesting ones, I sit down and write out my own thoughts about them, because AI still only has borrowed thoughts.
← Back to blog