How I run my own operations with AI
One ranked queue in place of a dozen separate agent pings. Agents do the legwork on a schedule, a chief of staff agent sorts it, and every call that needs judgment comes to me.
- I had scheduled agents doing real work for me, and each one reported on its own. Decisions piled up in lists nobody acted on.
- I set up a chief of staff agent over them and one ranked queue I clear from my phone, with rules I set and a log of every call.
- On the first night I cleared 24 decisions from that one queue. The system now learns from my reasons and brings rule changes back for my approval.
More agents meant more noise, not fewer decisions
Running a market launch taught me that the work isn't the hard part. Deciding what matters today is. My agents were doing the work: scanning, drafting, checking, logging. But every one of them messaged me separately, on its own schedule, in its own format.
The cost showed up as items nobody decided on. One list I checked had 59 entries waiting without a single call made on any of them.
Agents do the legwork. I make the calls.
- 1Agents work on a scheduleEach one owns a lane and a time slot, and follows written rules for what it may and may not do.
- 2Everything lands in one logEvery run records what it did, what it found, and anything that failed, with a checkpoint so a crashed run picks up where it stopped.
- 3A chief of staff agent builds the queueTwice a day it pulls what needs my judgment, removes duplicates and ranks it by deadline and money at stake.
- 4I decideApprove, approve with edits, reject or snooze, each with a reason. From my phone, in batches.
- 5The chief carries out my callInside the rules, and logged. Anything that would message someone, pay for something or post publicly stays with me.
Decisions and the options I turned down
One queue for everything
Turned down: a dashboard for each area. More places to check is the problem I was solving, not the answer.
Roll it out one agent at a time
Turned down: rewiring every agent the night it launched. The chief reads what the agents already log, and each agent switches over only after the one before it runs clean. Same way I'd open a new market: no big bang on a system that's working.
Who did what
Claude
- Ran each agent's work on schedule
- Built the queue page and the run logs to my spec
- Drafted rules and fixes from my decisions
Me
- Designed the system: lanes, chief of staff, one queue
- Set every rule, limit and what stays off limits
- Decide every judgment call and approve every rule change
- Capped usage at a set number of runs a week, reported every Friday
Early numbers, straight from the logs
AS OF NOV 8, 2026 · FIRST FULL WEEK OF NUMBERS POSTS NOV 15
A run came back short, and said so
On day one, a research agent finished well short of its target because many of the sites it checked wouldn't load. It logged itself as partial with the reason, instead of failing quietly or padding the results.
That night the review turned it into a standing rule: check the most reliable public pages first, then the sites that often fail. A small fix, but it's the habit that matters. Every miss becomes a rule, and any rule that changes what an agent does comes to me first.
The same pattern on a fleet
My own queue holds private data, so I built a version you can click through: the same rules, ranking and approvals, pointed at a 150 vehicle fleet with made up data.