Ask HN: What are agent sandboxes missing?

I'm building an agent sandbox platform with opinionated customizable templates for general purpose agents using the OpenCode SDK.

The idea is to provide a great CLI experience so it's easy to use an AI client like Claude Code or Codex to launch sandboxes, fetch/retrieve data and prompts, review generated outputs, and manage queues for ask-user questions.

I also plan to add an AI gateway layer, similar to Tailscale Aperture(not affiliated but one of their engineers gave a great talk at AI Engineer it's on YT) for monitoring network access, distributing secrets, and controlling what services/systems a sandbox has access to.

For people building agent workflows, I'd love to hear feedback if you've tried other sandbox solutions.

What was missing and what was the hard part?