Blog.

Writing on engineering, architecture, and lessons learned while building products.

Building an Agent Sandbox from Scratch (Part 2)

Part 2 of the sandbox series: observability, plan mode, self-repair, browser verification, and the runtime changes that made the system trustworthy.

Part 1 of the sandbox series: how I built a disposable runtime for AI agents with Docker-in-Docker, headless execution, and durable outputs.

The stack behind this blog — MDX, Shiki, Sandpack playgrounds, and the authoring pipeline that ties it all together.