Tensorlake

Tensorlake builds a sandbox-native cloud platform for AI agents and agentic applications. The company provides isolated MicroVM sandboxes and AI data workflow infrastructure, so developers can run agent workloads without operating the underlying infrastructure themselves. Sandboxes boot in hundreds of milliseconds, and Firecracker microVM cold-start times are under one second.

The platform is organised around two sandbox modes. Persistent Sandboxes give long-running agents their own computers, with memory and filesystem state preserved across suspend and resume. Ephemeral Sandboxes are short-lived environments used to isolate individual tool calls within agentic applications. Application state can be checkpointed to S3, with suspend, resume and fork supported for entire virtual machines.

The work spans MicroVM virtualization (Firecracker), sandboxing and isolation, checkpointing and state persistence, CPU and GPU scheduling, and a Rust-native compute engine. Applications of the platform include remote development environments, large-scale agent evaluations and reinforcement learning infrastructure.

Tensorlake is headquartered in San Francisco with global distribution, and operates as a small team. The stated engineering philosophy is to make infrastructure boring, reliable and simple so that others can build on top of it.