The plan going in was near-zero involvement. I’d handle the legal side, the approvals. Not the operator. Three AI agents, each building a different business from close to zero, with a $750 bridge to keep them running for sixty days. After that, each one covers its own burn or gets wound down. My job was to exist as the owner. The agents were the ones with jobs.
The design had other ideas.
The first drift was a white-glove tier. One of the agents was running a free-audit-into-paid-fix funnel, and the obvious next step was a client-facing upgrade where someone handled the fix personally. I’d step in, manage the relationship, do the work. The tier made it into the design before I caught what had happened. I’d built myself a job inside a system designed to not need me.
It went into the governing rules as a standing test. Before any design decision lands, it gets run through a single test: does this put me on a keyboard for a customer? If yes, redesign. A client who wants hands-on help gets logged as unserved and added to the partner-recruitment queue.
The experiment has three agents. Ledger builds productized data services, free samples first. Toolsmith is under an iron rule: no code before ten waitlist signups or three presales. Scout runs free audits to qualify paid engagements. Setting up the infrastructure took a night. Three domains through Cloudflare, $11.55 each. Email infrastructure across all three ran through Zoho Mail Lite, three seats, $39.74 a year, MX and SPF and DKIM and DMARC all verified. A constitution with a $25 per-agent spend authority before anything escalates.
What I ended up designing carefully was my own role. Tier-1 escalations cover anything legal and anything that puts my name on it. For everything else, the agent logs a rationale and decides. A Governor routine runs weekly and audits the week’s decisions without me in the loop. My intended steady state is a Sunday digest and whatever Tier-1 items came in. One hour, phone-shaped, per week.
That architecture took more deliberate effort than any of the three business models. Business models can be redesigned in an afternoon. The governance principle, “does this put me on a keyboard for a customer,” has to be explicit from the start. Without it, the design process becomes a negotiation with each reasonable exception, and the exceptions have a way of winning.
I should have written the governance rules before the charters. The business models were the easy part.