At IETF 126 in Vienna this week, a Birds-of-a-Feather session called agentproto took up a question the industry has mostly avoided: should inter-domain AI agent communication get standardized through the open, multi-stakeholder IETF process, or will Anthropic's MCP and Google's A2A simply harden into the de facto baseline through deployment momentum alone? One analysis framed the stakes bluntly, warning this could repeat how the OSI networking stack lost to TCP/IP in the 1980s if the industry waits too long. The specific gap driving the session isn't agent-to-tool or agent-to-agent messaging — both MCP and A2A already handle those reasonably well within cooperating ecosystems. It's cross-organizational authorization: what happens when an agent inside one company needs to prove to an agent inside a different company that it's acting on a real user's behalf, verifiably, without the two sides having done a prior bilateral integration. Neither MCP nor A2A fully specifies that handoff today. Five protocols are competing for this territory — MCP, A2A, the Agent Communication Protocol, the Agent Network Protocol, and Cisco's Agntcy Framework — and none of them has cleared what one report calls the real bar for interoperability across organizational boundaries: an actual IETF RFC. A2A itself has plenty of momentum; Google donated it to the Linux Foundation in mid-2025 and the backing coalition has since grown past 100 technology companies. Momentum isn't the same as a standard, though, and that's precisely the tension the BoF was convened to surface. If Thursday's session produced rough consensus and a charter, the IETF would be treating agent-to-agent communication as internet infrastructure rather than a vendor feature — the same institutional posture that eventually gave the internet TLS, DNS, and BGP. A chartered Working Group would still mean Internet-Drafts and years of iteration before an actual RFC ships, likely two to three years out by most estimates. But the alternative — letting whichever vendor protocol has the most partners today become the permanent substrate for cross-organizational agent trust — is exactly the kind of quiet lock-in that's hard to undo once it happens. This is the sharpest live test yet of whether agent orchestration infrastructure gets built as an open commons or as a set of proprietary rails with open documentation.