AWS Lambda MicroVMs offer isolated sandboxes for AI-generated code

Full analysis

AWS has launched Lambda MicroVMs, a serverless compute primitive designed to safely execute code generated by users or AI agents. Each MicroVM provides VM-level isolation (hardware-separated execution environments), near-instant launch and resume speeds, and state retention for up to 8 hours — addressing a core challenge for agent builders who need to run untrusted or AI-generated code without risking cross-tenant contamination. The feature is built on Firecracker, the same virtualization technology already powering over 15 trillion monthly Lambda function invocations.

The announcement is directly relevant to agentic systems that execute tool-generated code, such as coding assistants, data analytics platforms, and vulnerability scanners. Developers can create a MicroVM image from a Dockerfile, assign each user or agent session its own isolated VM with a dedicated HTTPS endpoint, and connect via HTTP/2, gRPC, or WebSockets. AWS explicitly highlights compatibility with its Agent Toolkit, signaling this as a first-class primitive for agentic workloads. It is available now in five AWS regions.