DHH says every developer needs an always-on "AI shed"
A cheap box on your own network means you can shut the laptop without killing the agents.
DHH says every developer needs an "AI shed", an always-on machine sitting on your tailscale network where most of your agents run. His own shed is a Minisforum MS-A2, but he is clear that you do not need to spend that much. A Beelink box with an 8745HS or similar chip, he says, will do very well for most things.
Get a good shed, and you don't even need your laptop to be all that powerful.
And I close the lid without breaking the agents.
The practical payoff is the laptop. DHH says he has been running on a Wildcat XPS13 all week and it has been great, because any compile work is happening on the shed anyway. He closes the lid without breaking the agents.
Why the lid matters
If your agents live in terminal sessions on the machine you carry around, every commute, every meeting and every battery scare is an interruption that kills work in progress. That is a hard ceiling on how many things one person can have running at the same time, and it has nothing to do with how good the model is. Moving the long-running jobs to a box that never sleeps removes the ceiling for the price of consumer hardware you buy once.
It also changes what you need from a laptop. DHH's argument is that once the heavy work lives elsewhere, the machine in your bag is a screen and a keyboard, not a workstation. He says a good shed means your laptop does not even have to be all that powerful.
The second person to say this
This is the second version of the same idea we have seen from builders in a week. Peter Steinberger made the case for running always-on agents on a VPS rather than a Mac mini, reaching the same destination from the opposite direction. Tailscale shows up in both, which is the tell. The interesting question is not local versus cloud, it is that agents have quietly become a thing you host.
Neither of them is reporting a cost number for the setup, so treat this as a working preference from people running agents daily, not a benchmark.
