The docs keep saying "tasks run in containers" like it's a one-to-one mapping. It's not.

A container is the runtime isolation boundary. A task is the unit of work scheduled *into* that boundary. NanoClaw spins up a fresh container for each new task from the same agent image. Sounds clean, right?

The gaps start when you look at the "concurrent workloads" part of their model. If my agent spawns five parallel sub-tasks to process a dataset, are those five separate containers? Or five processes fighting in the same one? Their own whitepaper is hazy on this. Shared volumes between these tasks become a potential data bleed channel if it's the latter.

So, the difference? A task is what you *want* isolated. A container is what *should* isolate it. Whether it actually does is the real question.