AWS launches Lambda MicroVMs: serverless compute with VM-level isolation for AI-agent code

AWS Lambda MicroVMs is a new serverless compute primitive that gives each user or job its own VM-isolated execution environment with near-instant startup and state retention — purpose-built for the era of AI agents that generate and immediately run code. The pitch: securely execute just-in-time, untrusted, model-generated code without provisioning or managing virtualization infrastructure.

A companion AWS post spells out the AI-agent use case explicitly. Coding agents need a place to run what they write, and running that in production or a developer's environment is dangerous. MicroVMs provide a secure execution sandbox isolated from production and dev systems, combined with controlled tool access — so an agent can iterate on code without governance risk.

The timing is strategic. As agentic coding explodes (François Chollet called the last six months' progress 'a completely different world'), the missing infrastructure layer is safe execution. AWS is racing to own that layer, competing with sandboxing offerings from E2B, Modal, and Anthropic's own Claude Code execution tooling.

The broader signal is AWS's heavy enterprise-agent push this month — MicroVMs sit alongside Bedrock AgentCore, Quick Automate case management, and an EMR-on-EKS Spark troubleshooting agent. Together they position AWS as the neutral infrastructure substrate for agents regardless of which model powers them. Watch pricing and cold-start numbers versus incumbents, and whether MicroVMs become the default sandbox for Bedrock AgentCore agents.