Is there possibility of something like this? transparent upgrade to sandbox when needed

[spa5k]

RFC

Summary

I was wondering, through current architecture, can we somehow do the above - https://blog.cloudflare.com/project-think/#the-execution-ladder, check this link a bit.

So basically have a Agent with just-bash/cloudflare shell etc, so that agent can read the files if needed, do typical MCP etc calls, but if needs arise to make some code changes, or something, spin up a real sandbox, and do things there, pretty natively/transparently.

Its not a feature request, but more of a exploration, or if you got any suggestion on it? I mainly dont want to run expensive sandbox until needed, and when it runs, it should be almost transparent but controllable.

[spa5k]

basically transparent or easy promotion between virtual sandbox -> to real sandbox when needed.

or any other idea? the agent can let's say require sandbox rarely, but it should be there, maybe a subagent?