Baseten launches stateful Sandboxes for AI agents

Baseten announced Sandboxes, an execution environment for agents that run multi-step tasks over long periods. The company describes it as the first piece of an "agent cloud" that combines compute, storage, networking and identity.

Each sandbox runs in its own microVM with its own kernel and hardware-level isolation. Baseten says this isolates untrusted agent-generated code more strongly than container-based setups and prevents lateral movement.

The main feature is persistent state. When an agent disconnects, Baseten snapshots the filesystem, memory and running processes, and compute scales to zero while the sandbox sleeps. The company says a sandbox resumes or boots from standby in 25 milliseconds, which it calls "up to 5x faster than other sandboxing tools." It also says deployments can scale to thousands of sandboxes within seconds. Baseten did not publish the benchmarks behind these numbers, name the tools it compared against, or give pricing.

Baseten says this lets agents carry a project across sessions and wait for input from people without paying for idle compute.

Two more components are due "in the coming months":

• Agent Drive, a shared, distributed filesystem across sandboxes, built for many writes of small files such as code and markdown.
• Agent Proxy, which controls network access from outside the sandbox. It injects credentials into calls, such as payments through Stripe, so the agent never sees them.

Baseten says the stack will also include permissions, audit logging, firewalling and static IPs. Sandboxes join its existing products, including Dedicated Inference, Model APIs, Training, Loops and Grid.