AWS Lambda introduces MicroVMs for isolated execution of user and AI-generated code

AWS introduced Lambda MicroVMs, a new serverless primitive that gives each session VM-level isolation with no shared kernel between sessions. Key features include near-instant launch and resume, full lifecycle control, and state preservation for up to 8 hours — a notable departure from Lambda's traditional ephemeral, stateless model. The explicit target is securely running user-supplied or AI-generated code without teams having to manage virtualization infrastructure themselves.

The timing reflects where agentic AI is going: as agents increasingly write and execute code, sandboxing untrusted output safely becomes a core platform need. MicroVMs let developers spin up strongly isolated sandboxes that can hold state across a multi-step agent run (up to 8 hours), then resume — useful for long-running coding agents, code interpreters, and per-user execution environments. This puts AWS in direct competition with sandbox offerings from the likes of E2B, Modal, and Cloudflare's isolate-based compute.

The launch was part of a broad AWS AI/data wave this week that also included Bedrock AgentCore Payments, a SageMaker Data Agent with Snowflake querying, multi-tenant RAG with Verified Permissions, and the AgentCore harness. Together they sketch AWS's pitch that it can run the full agent stack — execution, payments, retrieval, and authorization — natively. Watch pricing versus persistent compute, cold-resume latency in practice, and whether the 8-hour state window is enough for the longest agent workflows.