Modal VM Sandboxes went generally available on October 1, 2026, the AI cloud infrastructure company announced on its official blog: full Linux virtual machines built for AI agents. Flip the sandbox runtime flag to runtime="vm" and an agent gets a real computer to work in — Docker, local databases, even tinkering with the Linux kernel — while the APIs, images, and usage-based pricing stay exactly the same.
One flag: from container to a real machine
VM Sandboxes keep the Modal Sandbox way of working: the same APIs, reusable modal.Image images, sub-second cold starts, burstable CPU and memory, and hundreds of thousands of concurrent sandboxes per cluster. What changes is underneath: real virtual machines instead of gVisor containers. Modal built a custom runtime on top of the Rust-based Cloud Hypervisor, with its own work across host filesystems, lazy image loading, memory bursting, and snapshotting.
Switching costs next to nothing: the gVisor sandbox remains the default runtime, and moving to a real machine is a one-flag change. The official guidance is blunt — reach for the VM only when you hit the walls of userspace: Docker, FUSE filesystems, or niche kernel features.
Why agents want a computer
Modal tells a turning point in the post: earlier this year, the team watched Ramp build an internal agent that drives the entire software development lifecycle on Modal Sandboxes, and it became obvious what agents actually want from their runtime.
Agents running in evaluation or production increasingly want to live inside something that looks like a real machine: Docker stacks, local databases and dev servers, graphical environments, mobile simulators, even the Linux kernel itself. At the same time, Modal didn't want to lose the ergonomics it built up in the container era — the exec and filesystem APIs, the burstable resource model — so it combined both: the power of a VM with the feel of a container.
Three examples already running
Before GA, VM Sandboxes spent months with early customers, who launched more than 20 million VMs. The blog names three.
Linear's Coding Sessions let users hand an issue to a coding agent without leaving Linear; behind each session sits a VM holding the user's full dev environment. Linear's engineers say the switch was a single flag and everything just worked — Docker finally runs the way it does on a normal Linux host.
Legal AI company Legora uses it for long-horizon agent evaluations: each sandbox boots the entire Legora app — Postgres, a DOCX editor, the agent's own code sandboxes — running for hours across thousands of documents. Previously, simulating full Docker inside containers meant networking and FUSE workarounds at every layer; with VMs, every one of those workarounds got deleted.
Frontier AI data lab Snorkel uses it for agent simulation: millions of simulations a month, where agents must pull off zero-downtime database migrations and hot-swapping services under load in near-real-world environments. In their words, for a simulation to be meaningful, the environment has to look like the real world the agent will actually face — and VM Sandboxes give every simulation a complete machine.
This is not just a settings upgrade
Put the pieces together and the weight of the announcement becomes clear.
First, the competitive focus of agent infrastructure is moving down the stack. When models are hard to tell apart, whoever gives agents a more production-like "computer" wins the harder tasks. This release fills the "environment fidelity" gap.
Second, Modal is assembling both halves of the map: Modal Clusters, which went GA not long ago, targets multi-node training; VM Sandboxes target agent runtime and evaluation — one is where models get built, the other is where models do the work.
Third, the decision gets simpler for developers: if your workload runs happily in the existing sandbox, don't touch it; the moment your agent starts demanding real-machine privileges like Docker or kernel modules, flip one flag — no platform change, no image rewrites.
There are limits, of course: on day one of GA, VM Sandboxes are deliberately "container-shaped" for drop-in replacement, and new primitives built on the VM programming model are still on the way. Modal itself says the next step is making VMs lighter and more elastic.
But the direction is clear: in the 2026 agent race, the second half won't be won on models alone, but on what the "computer" that agents live in looks like.
