At IETF 126 in Vienna last week, the agentproto Birds-of-a-Feather session held what may be the most consequential procedural vote yet for AI agent infrastructure: whether to charter a formal Working Group tasked with producing the first binding RFC for AI agent communication. A preceding session, DMSC, had spent the meeting mapping the problem space rather than proposing a fix. Its target was the AI agent gateway model — the architectural layer sitting between an orchestrating agent and the downstream agents or tools it invokes. The scope is broad by design. A gateway has to handle capability exposure, request forwarding, coordination and synchronization across parallel agent threads, policy control over who can authorize an action, observability for human audit, and secure inter-agent communication so a receiving agent can verify a sender's identity and message integrity. Rather than pick a winner among existing approaches, DMSC's job was to profile how mechanisms already defined elsewhere — OAuth, WIMSE, DNSOP, and the A2A and MCP specifications themselves — cover pieces of that problem, and to flag where nothing covers it at all. That framing traces back to a January 2026 IETF blog post that named five unsolved questions no current protocol fully answers: how agents discover each other's existence and capabilities, how they delegate tasks and credentials across organizational boundaries, how user context stays private as it passes between systems, how multimodal content is handled as conversational context accumulates, and how humans stay in the loop before an agent commits an irreversible action. The stakes are less about any single company's protocol and more about whether the IETF's standards process will govern agent-to-agent communication at all, or defer to protocols like A2A and MCP that are already running at scale in production — A2A alone reports tens of millions of downloads a month. Session materials and recordings are being posted to the IETF Datatracker as sessions conclude, and IETF 127 is already scheduled for November in San Francisco, which gives the working-group question — assuming it charters — its next real deadline. What makes this worth watching isn't the vote itself but what it signals: agent-to-agent coordination has grown large enough, and consequential enough, that the body responsible for TCP/IP and DNS is now treating task delegation and cross-agent authorization as internet-plumbing problems, not vendor features.