The Agent Office

I run several businesses on my own, with AI agents doing real work in each one. This is the system that keeps dozens of them productive without letting any of them go beyond what they’re allowed to do. I use it every day.

• Agents working in parallel will overwrite each other’s work and start doing things nobody asked for if nothing stops them.
• The dangerous actions are specific: spending money, sending email, and posting anything public.
• I needed every one of those to come to me for approval, no matter which agent was asking.

• Each business gets its own lane, with a written charter that says what it owns and where it’s allowed to make changes. One lane never edits another lane’s files.
• One coordinating agent routes work between lanes and keeps the records. It doesn’t build anything itself.
• Every decision is written to a log that can be added to but never edited, so any agent can look back and see why something was done.
• Permission is set for each type of action, not for each agent, in four levels: do it and log it, do it and tell me, stop and ask, or I do it myself.
• Commands get checked before they run, so the rules are enforced by the system instead of only written into the instructions.

• I can run several businesses on my own, and every risky action still comes to me first.
• The checks have stopped my own mistakes more than once.
• In one recent week, the records caught a marketing plan that was about to send cold email from the wrong domain, and it was stopped before anything went out.