Post

Idle AI agents cost money. So does rebuilding the same environment for every task. In LinuxInsider, Jack M. Germain covers how Project Lunchbox tackles both, on Kubernetes clusters you already run. www.linuxinsider.com/story/edera-...

linuxinsider.comEdera Brings AI Agent Sandboxing to KubernetesPause it, clone it, run it again. Project Lunchbox aims to give developers sandbox convenience while keeping AI workloads on infrastructure they control.

Edera

@edera.dev

Live snapshotting lets an agent waiting on human input save its full state, suspend, and resume when needed. Teams can also prepare an environment once and reuse it, instead of repeating resource-intensive setup for every task or RL run.

Edera

@edera.dev

Copy-on-write forking clones a running microVM into isolated copies that share memory until they change a page. Each copy is a zone, created with one hypervisor call. “We can fork thousands of zones very quickly with this design,” says Edera CTO @alex.zenla.io

20/10/26, 6:55 pm1 repost2 likes

Edera

@edera.dev

“At the point we are freezing and CoW, we are talking milliseconds, whereas Kubernetes frequently measures things in seconds.” Getting there took custom tooling for freezing and cloning memory state. Alex: “We are often optimizing Kubernetes itself to make it all possible.”

Edera

@edera.dev

“The cost model then changes because the time spent on a task is lower when state can be shared.” It all runs on your own Kubernetes cluster. Full story: www.linuxinsider.com/story/edera-...