I run three businesses: an AV production company, a consulting practice and a service business. An agent team handles the busywork across all three. Eight agents, each with one lane: an orchestrator, inbox and calendar across four accounts, research, a weekday digest at 7:30am, a news radar that feeds it, and one agent per business.
None of them can send, spend or delete anything without my OK. Everything else is built around that rule.
The rules they run on
Each agent has one lane, and work is routed by who owns it. The inbox agent doesn't do research and the research agent doesn't touch the calendar. When everything has an owner, nothing gets done twice and nothing falls between two agents.
An approval covers one exact action. Approving a draft email approves that email, with that text, to that person, and nothing after it. Money, client inboxes, publishing and deleting stay approve-first for good.
A task is Working, Blocked or Done, and there are no other states. Blocked ends in a question I can answer in one word. Done means the outcome is true and there's proof I can open.
There's a correction budget of three build-and-fix cycles. On the fourth, the agent stops and reports what works, what doesn't and what it would need, instead of trying forever and handing me a mess.
Freedom is earned one job at a time. An agent gets more room on a specific task after it has done that task right, repeatedly, where I could see it.
Why this matters for your team
Most of the fear around automation is about software doing something nobody approved. That's a design choice, not a property of software. When I build for a team, the same rule ships with it: anything that leaves your building waits for a person's yes.
On the water utility quality plans, the software builds the plan and a person still reviews every one. In my own businesses, drafts sit in a queue until I say yes. The software does the assembly, and a person owns the outcome.
