Comment by Sai-HN

I had the chance to review the technical design, and it’s clear this project takes a thoughtful and well-directed approach.

Rather than relying on the usual trade-off between Firecracker-style microVMs and syscall-level sandboxing like gVisor, ERA takes a different path—leveraging libkrun to deliver “lightweight yet VM-like” isolation without compromising the developer experience. That balance is genuinely impressive.

What stood out to me while exploring the repository:

・Deeper isolation than gVisor—no direct access to the kernel surface

・A clearly safer boundary than containers, without the overhead of Firecracker

・Practical issues with libkrun (buildah, krunvm, case-sensitive volumes) are addressed with care in the README and setup scripts

・Maintains ~200ms microVM startup, making it fast enough to integrate naturally into agent execution loops

・Local-first by design, yet flexible enough to support Cloudflare Workers for orchestration when needed

・Well-crafted recipes and examples that go beyond the basics and support real-world usage

This isn’t just about “running microVMs”—it’s about delivering a tool that developers can actually rely on. It feels far from a proof of concept; it’s something you’d want to keep in your toolbox.

Running AI agents safely on local machines is still an open challenge with no clear standard. In that context, ERA’s approach—seamlessly integrating microVMs into everyday development workflows—is both timely and valuable. I have deep respect for the thoughtful implementation and design philosophy behind it.