Skip to content
Personal Learning Notes

Agent Sandbox OSS

A Kubernetes-native runtime for long-running, data-heavy AI agents

Why I built it

Putting agents into production runs into three walls fast — isolating each agent invites out-of-memory kills, provisioning a pod on demand costs 40 seconds or more, and a node failure takes the in-flight state with it. This project is me taking those three apart one at a time.

What it does

Sub-100ms cold starts

A pool of pre-warmed pods that requests claim directly, skipping scheduling and image pull.

Persistent workspaces

Working directories live on PVCs, so state survives a crashed or evicted pod and work can resume.

Async scheduling

A Redis-backed queue, so traffic spikes wait in line instead of competing for resources.

Seven progressive labs

Starting from the naive implementation, each lab fixes one problem the previous one exposed — the architectural evolution stays in the repo.

Why seven labs

Handed the final architecture cold, half the design looks unnecessary — a warm pool, PVCs, a queue; each one reads as over-engineering on its own. Splitting it into seven steps, where each step only solves what the previous one exposed, makes it visible what forced every component into existence.

Current status

In progress. Interfaces still move, no releases yet. Everything is open — issues about the design are welcome.