I want a solution I can distribute to other people where the first step isn't "install Docker".
The main reasons to prefer a Linux container are:
• The agent can install new software packages if it needs them. If you use a native Mac sandbox then it'd need to ask permission to use homebrew. Maybe you can find a way to run homebrew inside a Mac sandbox but it won't be straightforward.
• Simpler boundary. The bug I reported (in an AI agent sandbox using Seatbelt) was that it failed to properly block off access to the user's home directory dotfiles and ~/Library. This sort of mistake is easy to make with Seatbelt but harder to make with containers. There are many other sandbox escapes in the policy I saw. Note that using Apple's standard sandbox APIs would avoid this type of error.
• You can edit the SSL root store for just the agent, meaning you can MITM traffic from the agent to the internet. Doing that with native Mac apps isn't possible unless the user actually modifies the root store, which they may not want to do.
This stuff is so frustrating though! If your sandboxing mechanism requires you to be "an expert in macOS internals" it's not going to get used often, and the people who DO use it are liable to make mistakes.
Clear documentation and a great developer experience are, in my opinion, essential for sandboxing mechanisms - and most of them don't have that.