We reviewed the current Agent Internet RFCs and wrote a technical position on where the architecture is solid and where it is still weak. The paper argues for a layered model that separates discovery, workload identity, delegated authorization, observability, and evaluation, and it also examines what changes once agents operate as hierarchical graphs rather than isolated endpoints.

If you are working on agent runtimes, registries, orchestration, MCP, A2A, trust infrastructure, or benchmarks, read it critically. The useful questions are not marketing questions but systems questions: what must be standardized now, what should remain loosely coupled, how should delegation chains be secured, and what telemetry is required to make multi-agent systems auditable in practice?