Returns a process-scoped agent ID that is generated once and reused for the
lifetime of the process. The semantic scope of agent_id is
“one agent-block execution”, which is coarser than run_id (per-call).
Both currently collapse to the same generated value in single-run
invocations, but the conceptual distinction is preserved so that future
deployments can evolve the two scopes independently (e.g. long-running
daemon vs. per-request).