A working log about running my businesses with software agents. Fourteen of them, on a ten-year-old ThinkPad under my desk, every night between half past midnight and two.
Why I am doing this
I wanted to see for myself whether pretty much everything can be automated — not as a claim on a slide, but as systems that either ran last night or did not. The cheapest way to test that honestly was to use myself as the subject: my own products, my own mailbox, my own writing, my own mistakes. So this is not a study of other people’s automation. I am my own experiment, and this site is the lab notebook.
The second reason is plainer: I wanted to learn what actually works and what does not. Almost everything useful I have found here is negative — the guardrail that never fired, the green test that proved nothing, the nightly job that reported success for nine days while producing nothing. Those answers hold wherever this kind of system gets built, and they are worth more to me than the successes. Writing them down in public is what keeps me honest about them.
Who this is for
Founders and operators who have wired an AI agent into something real and discovered that the demo was the easy part. If you have ever had a nightly job report success for nine days while producing nothing, this is written for you.
It is not for people looking for prompt templates. Everything here is one person’s setup, described in enough detail that you can judge whether it would work for yours.
The four strands
- What I automated — one system at a time: what it does, what it replaced, and whether building it paid.
- How it runs — the plumbing. Schedules, gates, guardrails, and the parts that decide whether any of it survives a week.
- What broke — failures with the measurement attached. This is the strand I would read first.
- What it costs — tokens, hours and money, against what the same work cost before.
The editorial policy, in three sentences
Every number here came out of a log on my own machine. Where I could not measure something, I write that instead of estimating it — an unmeasured claim dressed as a measurement is worse than an open question.
I publish the failures at the same rate as the successes, because a log that only contains successes is marketing with a timestamp on it.
What the agents are not allowed to do
They draft. They research, write, generate images, grade their own output and file it. They do not publish, do not send, do not post, do not open accounts and do not spend money. Every piece on this site passed through a human deciding to release it, and that human is me.
That is not caution for its own sake. It is where the bottleneck sits, and pretending otherwise would be the first dishonest thing on the site.
Elsewhere
The build gets posted as it happens on X: @georg_singer. Professional profile on LinkedIn. Quantum computing and AI go to realquantum.ai; when a piece there involved something worth explaining about the build, the build half lands here.
What this is not
It is not a business. This is a personal log, written in my own time about systems I built for myself. Nothing on it is for sale, no services are offered here, there is no newsletter and there is nothing to book. If a post saves you an afternoon, that is the entire return I am after. The address on the contact page is for questions and corrections.