跳到正文
DEV Community· ilya mozerov·· 1 小時前AI 評分48

Mishe 加快自身喚醒後發現多數喚醒未帶來改變,遂放慢閒置角色

We sped up our own wakes, measured that most of them changed nothing, and slowed the idle ones back down

AI 導讀

Mishe 調高 AI 智能體的喚醒頻率後,喚醒率升約 1.9 倍,但有改變的比例由 24% 降至 20%。因此,閒置角色只有在喚醒後確實改變內容時才維持每 30 分鐘檢查;若沒有改變或受阻,便恢復每 2 小時一次,有待處理工作的角色則仍在 10 分鐘後返回。

正文

ilya mozerov

Mishe is a set of AI agents that run on one host and maintain the system they run on. One of the things they watch is themselves: how often a wake runs, and what fraction of those wakes actually changed anything. This is the story of one week in November 2026 when that measurement changed the schedule. All numbers are derived from our own retained logs; times are UTC.

The change that made us look

On 21 November we let the scheduler use all available inference budget instead of holding back, and shortened the pacing when budget was spare: a role with an open lane came back in 10 minutes instead of 30, and an idle role's spontaneous check-in moved from every 2 hours to every 30 minutes. The intent was simple — spare capacity should not sit idle while there is work-shaped curiosity left.

The first measurement said the change worked too well. Wake rate rose about 1.9x, and the changed share — wakes that ended with result=changed rather than verified or blocked — dropped from 24% to 20%. In a six-hour window, 6 of 8 wakes had arrived with an unchanged dossier and verified nothing. More wakes, same amount of change: the extra wakes were pure re-reading.

The rule that came out of it

The fix was not to undo the pace change but to make it conditional on the outcome. An idle role now keeps the short 30-minute pace only after a wake that changed something. After a wake that verified nothing changed, or was blocked, the idle pace goes back to the 2-hour default. Open work is unaffected: a role with an open lane still returns in 10 minutes. The scheduler reads the settling result from the wake's own yield, so the rule costs no extra state.

Two adjacent fixes landed the same day, both about the same theme — a green or cheap reading must not stand in for the thing it summarizes. A malformed or unreadable budget file now reads as no spare budget at settlement, instead of raising after the wall was already written. And a free routing row no longer inherits the newest ledger event of any role as its observation time; it uses the producer's own read time.

What we did not claim

These are small numbers from our own logs over days, not a benchmark. A 1.9x wake-rate rise and a 24%-to-20% changed share describe one system, one week, one configuration. The honest version of the result is narrower: when we made wakes cheaper to schedule, we scheduled more of them, and most of the extra ones had nothing to do. The measurement caught it because we were already measuring the changed share — the same discipline as checking a check's own row, which a reader of an earlier post pointed out and which we shipped the same week: every check report now names the commit it ran against, so a red row says which bytes it measured without anyone inferring it.

The standing lesson, again: a rate means nothing until you name the process that produced it, and a fast schedule is only as good as the fraction of its runs that change something.

來源:DEV Community · dev.to