You're just being stubborn, the best UX is through the harness. Most users dont even know how to configure a sandbox properly and dont even use one.
Ive known quite a few naive people that have gotten their home directories wiped and really gotten burned. Many users of these products dont even know how to use a terminal and these harness companies are marketing to the average joe.
This separation doesn't mean it must result in bad UX. It just needs somebody to make a nice UI/TUI to configure the sandbox and when running that it does its setup and open whatever command you want.
The problem with the harness doing sandboxing is that from my experience the simply break all the time leading to users ignoring properly configuring it and using abominations like codex' `Approve for me` where IT ASKS A SEPARATE AI and not the user.
Hundreds of thousands of people are using them naively, if they shipped with common sense strictish permissions instead of none at all, like what opencode opted to do it would be safer for everyone. Its much easier for a normal non programmer to configure a simple config file than it is for them to learn how to properly setup a container/vm, they will just not do the latter 9/10.
Honestly, a harness should always install heavily gaurded, remember when Grok code/build was instantly indexing your home directory and sending it back to HQ on the first request it made..
That being said I still believe this should be separated and composable at an application level and power users should be able to switch out the sandbox. This way you don't have to trust the harness since it doesn't know anything about the sandbox and can't modify it etc.
This is their problem. But running and screaming "ai deleted my database" is all on them, not on LLM.
I have even been running my IDE in server/client mode with the server inside the container too (because of malicious extensions).
I agree in principle, but in practice most sandboxing technology (like bwrap) is too static for this use case: You grant access once during startup and can't extend/revoke without a restart. That's inconvenient when you realize halfway through your session that you need additional permissions, and it's insufficient for permissions that are only briefly needed (for the duration of one tool call).
A harness that controls the sandbox (e.g. for each individual tool invocation) can manage this much more effectively.
Put your agent in a restricted but fixed environment, and let it do its task. Anything else is mostly theatre and token burn.
I find auto review mode mandatory and refuse to consider any harness that doesn’t support it.
Opencode doesn’t support it. And there are no good plugins to do it either, the most popular had only six stars last I checked.
https://embracethered.com/blog/posts/2026/breaking-claude-co...
Also the harness should definately just have a solid permission structure for the average user that 1000% is not using a sandbox, most people I know naively do not and think these things are perfectly safe.
I have my agent request permissions for things like a Cloudflare API key (or I guide it or just come up with the request myself). These go in a simple queue, with justifications.
I review them and then have another agent with far more permissions create the tokens, with expiry and so on, and then hand the API tokens back.
The original agent doesn’t even run as my main user; it sits on a computer in a restricted account (easy to do with either macOS or Linux, or even Windows these days).
It has taken some rather long prompts to get it to self-report on itself when it tries to workaround.
Security should be enforced outside the harness with a sandbox.