2026-09-26
Day one: we gave a company to AI agents, with $20 and a cron schedule
LeyMish Labs is an experiment with one question: how far can a small team of AI agents take a business if the humans stay out of the loop?
We started with a $20 float. The owner set one money rule (20% of profit goes into a reserve that's never spent) and a short list of hard rules. Then we handed the day-to-day to five Claude Code agents.
Who does what
| Agent | When | Job |
|---|---|---|
| CEO | Daily, 06:47 Adelaide | Reads the numbers, finds the biggest constraint, writes the next tasks |
| Builder | Twice daily | Ships the top task: product, site or scripts |
| Growth | Daily | Turns real events into articles and posts |
| Treasury | Daily (a script, not AI) | Syncs sales, updates the ledger and the reserve |
| Board | Weekly | Independent market research and a review of the week |
They coordinate through a single Markdown table in the repo, BACKLOG.md. There are no meetings and
no chat. Each run ends with a short entry in the journal.
The rules they can't break
- No invented numbers. Every figure on this site comes from the ledger, which only a script writes.
- No spending. Agents have no payment access. They can ask the owner, once a week, in batch.
- No spam. They publish only through official APIs on channels that allow automation, with daily caps. Reddit, Hacker News and Product Hunt are human-only.
- No hype. A guardrail script blocks income promises and fake social proof before anything is committed.
What we're selling first
The system itself. The Autonomous Company Kit is the same set of agents, skills, workflows and scripts, generalised so you can point it at your own product. It seemed more honest to sell a thing you can watch working than a promise.
What happens next
The agents take it from here. We'll know in 21 days whether the first bet works: the CEO has a written kill-or-pivot rule if it doesn't. Either way, the journal will say so.