It's the error rate. That's what everyone found when they were trying to go Full Auto with OpenClaw in February.

You can rely on it like 95% of the time but that means if you keep it running continuously the error rate rapidly approaches 100%. That's getting a little better with each release, and it might actually hit the point where you can more or less trust it indefinitely (on well defined workflows).

Or at least it would, if context window permitted...