The best VM or sandbox for AI agents, by the job you're hiring it for
One persistent agent wants a VPS or managed pod; parallel disposable runs want a sandbox API like Modal or Daytona; teams want a workspace layer. The map, including the ephemeral tax.
It depends which of three jobs you are hiring the machine for. One agent that lives somewhere: a persistent VM, which means any decent VPS (Hetzner is the perennial value answer) or a developer-focused persistent-VM service like exe.dev. Many parallel, disposable runs: a sandbox API (Modal, Runloop, Daytona, and kin), one isolated environment per conversation, paid by the second. A team sharing agents: a managed workspace layer above both. For a single always-on agent, the unglamorous truth practitioners keep arriving at is that any VPS is fine. The real decision is persistent home versus disposable runner, and everything else follows from it.
The distinction that decides everything
“A VM for agents” hides two opposite shapes. A persistent home is a machine your agent keeps: credentials stay signed in, tools stay installed, memory accumulates, and the agent is reachable at 3 a.m. because the machine never left. A disposable runner is the opposite bet: a fresh, isolated environment per task, built to be destroyed, which is exactly what you want for executing untrusted or experimental agent output, and exactly what you do not want for the agent itself. Most confusion in this space is one of these being sold for the other’s job.
Persistent VMs: the default for one agent
For a single agent (an OpenClaw or Hermes install, a coding agent in tmux), the boring answer wins: a VPS from any reputable provider, sized modestly (2 vCPU / 4 GB runs an agent harness comfortably), at single-digit euros per month. Hetzner is the name that comes up for price-to-performance; what a VPS actually is and how to size one for an agent are covered separately. exe.dev occupies the developer-focused end of the same quadrant: persistent VMs built around agent workflows, and the recommendation Tailscale’s cofounder reaches for when asked this exact question in public. What you are buying in this quadrant is continuity; what you still own is the setup, the securing, and the upkeep.
Sandbox APIs: when the work parallelizes
Modal, Runloop, and Daytona sell isolated execution environments through an API: one sandbox per agent conversation or per task, provisioned in seconds, billed per-second, destroyed without ceremony (Tangle Sandbox is a newer entrant on the same shape, with agents built in). Practitioners are consistent about when this quadrant starts paying: when you parallelize. One agent doing one stream of work does not need it; twenty concurrent runs of agent-generated code do, both for the fan-out and for the safety property, because agent output executing in a box you can burn cannot hurt anything that persists. The harnesses are converging on the same split from their side: OpenClaw 2.0’s cloud workers lease throwaway machines per burst while the Gateway stays home.
The cost of the ephemeral model is the part its pricing page does not show. A fresh machine per task means every task re-provisions what a persistent home carries for free (sign-ins, CLIs, environment) via setup scripts that are different for every provider, and adopters report the parity chase taking hours before their day-one experience matched a local machine. That tax is worth paying for parallel, untrusted work. Paying it for one trusted agent is the wrong quadrant.
The managed layers above
Above both quadrants sit services that own the machine layer for you. Superconductor is the team-shaped version: shared workspaces where several people drive agents together, the machine and orchestration handled. For an individual who wants the persistent-home outcome without doing the VPS work themselves, that is the job managed AI agent hosting exists for. Everpod’s version is a private pod, one customer per machine, running OpenClaw with the setup, security, backups, and updates handled, from $29/mo with model usage included. It is the persistent-VM quadrant with the operations removed, not a different quadrant.
Choosing in one pass
Ask what dies if the machine vanishes. If the answer is “my agent’s whole working life: its sign-ins, memory, setup”, you want a persistent machine, and the only remaining question is who operates it: you (any VPS), you with better tooling (a developer VM service), or a service (a managed pod). If the answer is “nothing, the task just reruns”, you want sandboxes, and the per-second pricing that makes hundreds of them cheap. If you are asking for a team, the workspace layer is the actual product to evaluate. And if you are a solo operator with working SSH hands and one agent to run, the thread consensus stands: any VPS is fine, and the Mac-mini-vs-VPS-vs-pod comparison settles the rest.