Abstract

A method is disclosed for triaging an alarm raised by a cumulative-sum change detector over the model-output tokens consumed by an autonomous software agent, to tell a run that keeps re-executing without converging from a burst of parallel, bounded work. A periodic sentinel builds a zero-filled series of fixed-width time buckets per agent, measures it against the agent's own robust baseline (a median and an always-positive scale derived from the median absolute deviation), runs a one-sided upper cumulative sum, and on an alarm assembles the evidence, including per-unit-of-work token totals, into one object that is triaged and, whenever a full-severity alarm record is written, persisted with it. A pure classifier reads that object and returns likely-runaway, inconclusive or likely-legitimate. Because a time bucket sums every session active in it, the classifier reconciles two grains: an extreme bucket peak is withheld as runaway evidence only when per-unit window totals show enough distinct units, none reaching runaway magnitude (so the peak must combine several units) and none dominant; unless a sustained-streak arm fires, a unit-grain resolver then decides on the largest unit or leaves the verdict inconclusive. A peak-only runaway verdict whose trailing buckets have gone quiet is softened, but only to inconclusive: the softening never yields likely-legitimate, never applies to a sustained-streak verdict and bypasses the unit-grain resolver, so it can never trigger the triage verdict's only routing effect, a severity reduction permitted on likely-legitimate. Absent or unreadable per-unit evidence leaves the bucket-grain verdict standing.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS