Most of the thread is on the approve-versus-act threshold, so a different always-on failure mode: standing intent going stale. If I tell Spark on Monday to watch for something and act, and my situation changes by Thursday, does it re-check the instruction is still wanted before firing, or act on the original intent? On-demand agents sidestep this because you restate context every time you prompt. Persistent ones quietly accumulate stale standing orders, and that feels like where the awkward actions would actually come from.
我很好奇Gemini Spark如何决定何时请求批准、何时自主行动。是基于任务类型、用户自定义规则，还是它会随时间学习偏好？