[simonw]

I think sandboxes are having their moment because it's become undeniable that coding agents are useful, and that they're more useful if you run them in YOLO mode rather than having to approve everything they want to do.

Coding agents are still a relatively new category to most people. Claude Code dates back to February last year, and it took a while for the general engineering public to understand why that format - coding LLMs that can execute and iterate on the code they are writing - was such a big deal.

As a result the demand for good sandboxing options is skyrocketing.

It also takes a while for new solutions to spin up - if someone realized sandboxes were a good commercial idea back in September last year the products they built may only just be ready for people to start trying out today.

[ambicapter]

Why/how are they more useful in YOLO mode than in careful mode?

[simonw]

You can literally give them a task that will take a couple of hours to finish (like "port this library to language X, start by porting the tests, don't stop until all of the tests pass against the new implementation"), go out for lunch, come back and they'll have finished and probably got it ~90% right.

[phainopepla2]

It's just a lot easier to let them run loose and finish a task before reviewing it, rather than have to babysit and approve every command they want to run. It frees you up to do other things in that time. For some people, that's running more agents in a different terminal, for others that doing something else entirely.

[esperent]

The flow I'm using is plan -> technical plan -> execute using TDD.

My level of involvement decreases from step to step. I'm totally in control of the initial plan. I'm giving strong oversight of the technical plan. But by the time it comes to executing, I'm happy to let it completely take over and I'll review either at the end, or break it down into 2 - 4 phases for long plans and I'll review after each phases.

For this final step, which might be 30 minutes, I'll step out and do something else. I want to be sure nothing bad will happen on my machine if I do that, so sandboxing is important.

[skinner927]

Look up Ralph

[theblazehen]

To expand - This refers to the Ralph Wiggum loop, which keeps repeating a prompt to the agent until it responds with a completion promise

https://awesomeclaude.ai/ralph-wiggum has some tips and examples of it