Running a fleet of agents from a phone
Tuesday, September 29, 2026
You don't need a desk to run agents. Two old machines stay on, the agents run on them, and a phone is where work goes in and approvals come out. The person's job shrinks to the steps that can't be undone.
I do most of my work away from a desk now.
That isn't a productivity trick. It's what happens when the work runs somewhere else.
The setup
Two old machines that stay on. A Mac and a Linux laptop that would otherwise be in a drawer.
Agents run on them, in sessions that restart themselves after a crash or a reboot. A watchdog checks that they're alive.
The phone is where work goes in and approvals come out.
What runs on it
Everything on the portfolio.
closet.ai re-reads six fashion retailers every morning. The run takes about seven hours, and nobody watches it.
A directory of Lebanese businesses went from 1,622 listings to 4,569 in a day, built from public data sources.
Lurker, a tool that finds threads worth replying to, went from first commit to live site in two days.
Across twenty repositories, that was about 800 commits in thirty days.
What I do
I say what to build and what done looks like.
I design the checks. A check is what lets me not read the code.
And I approve the steps that can't be undone: merges, deploys, anything that spends money.
Everything else, I read about later.
What breaks
Machines reboot. A keyring locks after a restart, and every push fails until it is unlocked. A scheduled job stops firing without an error.
So the rule for the setup is the same as for the agents: fail loudly. A job that silently does nothing is worse than one that crashes.
The schedule that stopped firing was replaced by a wake every five minutes that checks whether it's time. Less precise. Never silent.