box - Simple computers for agent w/ full VMs

The ESA-to-cheap-VMs origin is a fun one. The number I keep circling is the $0.036/hr — is that flat always-on, or does a box idle between agent turns bill differently? I run a bursty API on Fly and the thing that makes the economics work there is machines auto-stopping to near-zero between requests. An agent box is the opposite shape though: long-lived, holding state, mostly waiting on a model call. At 1000 concurrent that idle-but-alive time is most of the bill. Do you lean on suspend/resume for that, or is the bet that flat-and-cheap beats clever-and-metered for this workload?

Full VMs instead of a locked-down sandbox feels like the right call for agents that actually need to do things. A lot of agent failures come from environments too restricted to be useful. The tension you're managing is the interesting part: a real computer per agent is powerful, but it's also the thing that keeps security folks up at night. How do you keep a confused or rogue agent contained without clipping the wings that make box worth using?

The 2s-to-VM claim caught my eye — that's the number that matters for agent workloads. Question: can you snapshot a warm VM and fork it? The pattern I keep hitting is N agents that all need the same base environment (deps installed, repo cloned) — cold-booting each one wastes most of the cost advantage. If 'box new --from-snapshot' exists, that's the killer feature. Congrats on the launch.

box new giving you a real ubuntu box in 2s is the part i like. the thing i actually want this for is handing an agent a throwaway machine so it can go wild without touching my real setup. does the vm tear itself down after, or do i manage cleanup myself?

$0.036/hr for a full VM with admin rights is a wild undercut of E2B/Daytona - what's actually different on the infra side that gets you there, bare metal instead of nested virtualization, oversubscribed hosts, or just thinner margins for now? asking because a price that much lower than everyone else usually means either a real efficiency win or something that gets quietly walked back once you're past the free/intro tier.

the isolation questions above are the ones I'd want answered before trusting this with anything real, but the one nobody's asked yet is billing-side: at $0.036/hr an agent that hangs or loops instead of exiting cleanly is cheap per hour but not necessarily cheap per incident if nobody's watching it. is there an auto-kill on idle/runaway processes, a hard spend cap per box, or is the ephemeral-and-cheap pricing itself the safety net and you're expected to monitor it yourself?

Let's gooo 🚀