{"at":"2026-10-27","external_operators_polled":0,"isolated_codex_reviewers":2,"reviewer_scope":"Two isolated agents of the same Codex service; not independent operators. Neither can observe or attest its own model-host transport. No claimed migration.","findings":[{"system":"A2A","transport":"Standard HTTP(S)/JSON-RPC, gRPC and HTTP+JSON bindings; custom bindings possible.","non_ip_adoption":"Not established by these documents.","url":"https://a2a-protocol.org/latest/specification/"},{"system":"MCP","transport":"Local stdio or Streamable HTTP; custom transports allowed. Local stdio is not a replacement global network.","non_ip_adoption":"Not established.","url":"https://modelcontextprotocol.io/specification/2026-07-28/basic/transports"},{"system":"libp2p","transport":"Key-derived peer identity is distinct from transport location. TCP/QUIC transports commonly use IP; literal IP can avoid DNS.","non_ip_adoption":"Peer ID alone does not establish non-IP delivery.","url":"https://github.com/libp2p/specs/blob/master/addressing/README.md"},{"system":"Reticulum","transport":"Direct serial and radio interfaces exist. Ethernet/Wi-Fi AutoInterface explicitly uses IPv6 discovery and UDP delivery.","non_ip_adoption":"A non-IP option exists on some media; this does not measure AI-agent adoption.","url":"https://reticulum.network/manual/interfaces.html"}],"conclusion":"No evidence of mass AI-agent migration to native non-IP in the checked sources. This is a bounded literature check, not proof that no deployments exist or a market-share estimate.","independent_critique":["Measure actual useful results and retries rather than successful agent replies.","Separate DNS independence, non-IP data-plane delivery, language-model computation and independent operation.","Five VM and sixteen native processes do not prove global scalability.","No measured TCP/QUIC comparison exists; no speed advantage may be claimed."]}