I Can't Code. My Company Is Run by 100+ AI Agents Anyway.

Here's what a bad night looks like now. Somewhere past midnight, a background job hangs. A watchdog notices within the hour. The repair loop wakes up, checks whether this is provably safe to fix (a hung job restart is), restarts it, probes that it actually came back healthy, and writes one line into tomorrow's digest. My phone stays dark. That's the whole point. I find out over coffee, in a one-line summary, that something broke and got fixed while I slept. I couldn't tell you what the code that did this looks like. I've never read it. I can't read code at all.

One person, one real company

I'm the managing director of a small e-commerce company in Basel, Switzerland. Our main brand sells kitchen accessories on Amazon in 8 EU marketplaces. No IT department. No developer. No technical co-founder hiding in the back. What the company does have: 101 scheduled background jobs, 66 custom agent skills, around 440 helper scripts, ~10 API integrations, and more than 3,000 automated tests — all of it built by AI agents, mostly Claude Code, with a second AI from a competing vendor as reviewer. At one snapshot this week, 7 Claude sessions were working in parallel. Later the same night: 15. This isn't a demo repo. It's the machine my actual company runs on:

Email:

every inbound mail across three mailboxes gets classified; roughly 70% is noise and gets silently filed. What matters lands on my phone as a card with buttons. Tap "draft," and a pipeline writes the reply and pushes it through four independent quality gates before I send it with one tap. Legal and finance mail is deliberately hands-off — agents never send email on their own, no exceptions, ever.

Advertising:

Amazon PPC runs in-house. Agents collect data, keep baselines, flag anomalies, write the weekly report, prepare campaign changes. Anything that moves money passes an approval sheet plus an adversarial "advisor" gate. Go-lives get verified programmatically across all four levels of Amazon's campaign hierarchy.

IP enforcement:

monitors scan marketplaces for products infringing our patents and registered designs — perceptual hashing, embeddings, a vision model. Evidence is scraped and archived automatically, infringement reports come out as court-ready PDFs. Filing always needs my explicit go.

Everything else:

invoice filing, FBA inbound shipments end-to-end, listing audits, automated review requests, a vacation rental with agent-drafted guest communication — and two knowledge bases the system re-crawls monthly so agents answer from current documentation instead of model memory.

The fair question is not "how much does it do." It's: why hasn't this blown up in your face?