Wouldn't you want to be in control of your dependencies in this case anyway? Like why would you ever want it to autodownload sdkmanager? Doesn't this seem like bad idea?

I think the architectural mistake is letting the agent execute the environment setup instead of just defining it. If you constrain the agent to generating a Dockerfile, you get a deterministic artifact a human can review before any license is actually accepted. It also saves a significant amount of tokens since you don't need to feed the verbose shell output back into the context loop.

For something you throw away next week, just need something simple and you run everything isolated? Why not?