Does OpenTelemetry identify the top-level agent in a multi-agent system?

Yes, as of a convention merged September 30, 2026: gen_ai.main_agent is a resource-level entity, identified by gen_ai.main_agent.id, that names the one persistent, top-level agent running in a process, something with a stable identifier across invocations like a scheduled job, a hosted agent in a cloud console, or an agent declared in an A2A Agent Card.

It’s deliberately narrow. The entity must not be emitted at all if the agent has no stable identifier, a transient in-memory id doesn’t qualify, and it isn’t meant for telling a main agent apart from the subagents it calls inside one run; that distinction already belongs to the gen_ai.agent.* span attributes on each invoke_agent operation. Grading each agent in a multi-agent run separately already misses the failures that only show up between them, and a resource-level id for the outer agent doesn’t change that: it says which deployed agent a trace came from, not how the agents inside that run behaved together. Like the rest of the GenAI spec, it’s marked development stability and can still change.