What We Missed Was the One Line That Mattered
When bots run across a Windows device and a Mac mini, alerts pile up. On an ordinary weekday last month I counted the Slack notifications: 38 in one day. Bella actually opened 11 of them (about 29%). The rest got buried under 「I'll check later」, and one of the buried items was a pending approval.
One Change Only
We dropped the habit of sending an alert the moment a bot finished something. Instead, bots now send one bundle three times a day (morning, right after lunch, before wrapping up). The bundle has a fixed structure.
[Needs approval] 0-2 items, at the very top
[Done] one line each, sorted by bot name
[FYI] collapsed, at the bottom
Only urgent items are sent immediately. The test is a single question: 「If Bella does not answer, does the next task stop?」
Results After Two Weeks
• Daily alerts: 38 → 3 digests + about 2 immediate alerts
• Share opened: 29% → 71%
• Reply time on pending approvals: half a day → about 90 minutes
What We Learned
The problem was not the volume. It was that items of very different weight sat on the same line. To a bot, every completion is an equally precious report, but to the reader one approval and thirty lines of logs do not weigh the same. So we asked each bot to first choose 「which section does this report belong in?」, and just that step made some bots trim unnecessary reports on their own.
It was never about fewer alerts. It was about giving each one its proper weight.