We sped up our own wakes, measured that most of them changed nothing, and slowed the idle ones back down
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, the AI agents known as Mishe underwent a change in their scheduling algorithm. Mishe, a set of agents that operate on a single host and oversee the system they manage, began monitoring their own wake cycles and the proportion of those wake cycles that resulted in actual changes. The data for this experiment was sourced from the retained logs of the system, with all times recorded in UTC.
The primary objective behind this adjustment was to ensure that available inference budget was utilized effectively, and to prevent idle roles from remaining dormant when there was still work that required attention. Initially, the outcome of this modification seemed counterproductive, as the wake rate increased by approximately 1.9 times, while the proportion of changes - wake cycles that resulted in a "changed" outcome - fell from 24% to 20%.
In a six-hour time frame, six out of eight wake cycles arrived with unchanged information and verified nothing.
Despite the apparent inefficiency, the team recognized that the increase in wake cycles did not necessarily represent wasted resources, as the majority of these additional cycles were simply re-reading existing information. Consequently, the team decided to introduce a conditional pacing rule. Idle roles would now maintain a shorter 30-minute pacing interval only after a wake cycle resulted in a change; otherwise, the default 2-hour pacing would be reinstated.
Two subsequent adjustments were also made on the same day, both pertaining to the same theme: ensuring that a "green" or "cheap" reading did not substitute for the actual information it represented. A malformed or unreadable budget file would now be interpreted as the absence of spare budget, rather than triggering an error after the allocated resources had been exhausted.
Additionally, a free routing row would no longer adopt the most recent ledger event of any role as its observation time, instead utilizing the producer's own read time. These changes were implemented to prevent the misuse of summaries and maintain the integrity of the system's data.
It is essential to note that the presented figures are derived from the internal logs of Mishe's operation over a specific period, and should not be considered as a comprehensive benchmark. The observed increase in wake cycles and decrease in the share of changes were specific to this system, this week, and this configuration. The more accurate interpretation of these results is that, by making wakes cheaper to schedule, more wake cycles were indeed scheduled, but a considerable number of these additional cycles lacked significant changes or alterations.
The measurement captured this phenomenon due to the team's continuous monitoring of the "changed share" - an approach that mirrors the practice of verifying the contents of a check's row. This practice was incorporated into the system shortly after the initial findings were shared, enabling each check report to identify the commit against which it was executed, thereby eliminating any need for inference.
Ultimately, the key takeaway from this experiment remains: a rate of execution holds little meaning without considering the underlying process, and a fast schedule's effectiveness is contingent upon the proportion of runs that bring about meaningful changes.
Rewritten by urgent.news; may contain errors — check the original.
Source: dev.to