TL;DR: An open-source microVM orchestrator that runs untrusted agent code in secure, hardware-isolated environments with sub-second boot times.
Summary: Celesto AI released SmolVM, an open-source tool that provides hardware-isolated microVM sandboxes for AI agents. Utilizing Firecracker on Linux and QEMU on macOS, it allows agents to run terminal commands, execute untrusted code, and browse the web safely. It boots in under 500ms and supports state snapshotting, file sharing, and persistence across agent sessions.
Why it matters: AI developers can execute agent-generated code safely without exposing host infrastructure or relying on heavy container setups. Integrate SmolVM's Python SDK into your agent framework to run untrusted code blocks securely.
Source: @tom_doerr