Why AI coding agents need a persistent machine

Ephemeral was the right answer to a different question

Codespaces, ephemeral CI runners, per-task sandboxes — the design center of the last generation of cloud dev tooling is disposability. Spin up clean, do the thing, throw it away. That was the right answer when the expensive resource was human attention and the failure mode was "works on my machine."

It assumed something about the workload, though: that sessions are short, that state between sessions is a liability, and that the human closes the laptop when they're done because the work stops when they do.

Agents break every one of those assumptions

An agentic CLI run is a long-lived process. It plans, edits, builds, runs tests, reads the failures, and loops — for minutes or hours, unattended if you let it. Three properties follow:

• The process must outlive your attention. The whole point of autonomy is that you leave. If leaving kills the run, you don't have autonomy — you have autocomplete with extra steps.
• State is an asset, not a liability. The half-finished branch, the running dev server, the node_modules cache, the agent's own session context — an agent resumes productive work in seconds on a warm machine and spends real time and real tokens rebuilding on a cold one.
• The failure domain should not be your pocket. A run that lives on your laptop inherits every hazard your laptop faces: sleep, battery, Wi-Fi, the commute, the OS update.

"Just use a sandbox" gets the isolation right and the lifecycle wrong

Per-task cloud sandboxes solve containment, and containment matters — an agent with shell access should absolutely be isolated from everything else you own. But most are containers on shared kernels with lifetimes measured in minutes, priced by the second, that discard state on exit. Leave one running around the clock — which is what an always-available agent means — and both the pricing model and the design intent fight you.

What the workload actually wants is closer to the oldest thing in the catalog: a machine. Persistent disk. Real isolation. Root. A lifecycle measured in months. The boring Unix answer, with modern isolation underneath.

The remaining question is who operates it

You can absolutely run this yourself on a VPS — here's the complete recipe, and for some people it's the right call. The trade is that you become the operator: provisioning, hardening, backups, restores, disk pressure, patching. None of it hard; all of it recurring; all of it competing with the work you bought the agent to accelerate.

That operating job is the part we think should be a product. DevRig is a persistent Linux machine for your agents — your agent CLI, your model account, a terminal that resumes exactly where you left it — with the operating included. We open in October; the waitlist locks launch pricing.