Modal makes VM Sandboxes generally available for AI agents

A full computer behind one flag

Modal says VM Sandboxes are built for teams that want to give agents the power of a full computer. Developers choose the new runtime by setting runtime="vm" when creating a modal.Sandbox. The company's example runs a command inside the VM and gets back a Linux kernel string.

According to Modal, the VM runtime brings the same APIs, modal.Images, sub-second cold-starts and CPU/memory bursting as before. It also supports the hundreds of thousands of concurrent Sandboxes that its users are accustomed to.

A second example in the announcement starts a Docker-in-Docker image, waits for a readiness probe, and then runs the hello-world container from inside the Sandbox.

Why Modal built it

Modal says it was originally built around serverless GPUs, using a gVisor-based runtime to scale lightweight GPU containers quickly. Its Sandbox product was also built on gVisor. Earlier this year, the company watched Ramp build an internal agent that drives the entire software development lifecycle on Modal Sandboxes, which made clear where agents were headed.

Modal argues that agents increasingly want to live inside something that looks like a real machine. It lists Docker stacks, local databases, dev servers, graphical environments, mobile simulators and work on the Linux Kernel itself. In the company's words, "agents just want a computer."

On the engineering side, Modal says it started from the Rust-based Cloud Hypervisor project and added a half dozen innovations across host filesystems, lazy image loading, memory bursting and snapshotting. It promises deep dives on this work in the coming weeks.

Early customers

Modal says early customers launched over 20 million VMs in the last couple of months. It highlights three:

• Linear: Every Linear Coding Session now runs on a VM Sandbox. Linear spins up the user's full dev environment in the VM and uses cgroups and network namespaces to keep the coding agent, the dev server and its own host process from competing for resources. Engineer Ryan Delaney said switching "took a single flag, and everything just worked."
• Legora: The legal agent company uses VM Sandboxes for evals, building its full app in each Sandbox, including Postgres, a DOCX editor and the agent's own code sandboxes. Eirik Drage Steen said Docker now behaves like it does on a normal Linux host and the team deleted its networking and FUSE workarounds.
• Snorkel: The AI data lab runs millions of agent simulations a month with Harbor, on scenarios such as zero-downtime database migrations and hot-swapping services under load. Rustem Feyzkhanov described the VM as a full machine for every simulation.

Should teams switch?

Modal's answer is that it depends. gVisor remains the default runtime, and workloads that run well on it today will keep doing so. The company recommends runtime="vm" when a workload runs into the limits of userspace, such as Docker, FUSE filesystems or niche Linux kernel features.

Both runtimes share the same APIs, Images and usage-based pricing, so moving between them is a matter of flipping a flag. Modal does not give price figures in the announcement.

Looking ahead, Modal says VM Sandboxes are intentionally container-shaped for now so they can be a drop-in runtime compatible with existing code. It plans to build new primitives that use the broader capabilities of VMs and to make VMs more elastic and lightweight.