From November 14 to 20, the Internet Engineering Task Force meets in San Francisco — its 127th full meeting. One of the working groups on the agenda is agentproto, which is writing the baseline protocol for how AI agents hold conversations across platforms and vendors. I wrote last week about the group’s charter and what it leaves out. The better question, and the one more people should be able to answer, is how this whole machine works. Who writes these rules, and how do you watch or push back?
Why it is called 127
The number is just a count. The IETF has met roughly three times a year since its first meeting in 1986, and each gathering gets a sequential number. There is nothing special about 127 — it is simply the 127th time this group has convened. The IETF is not a company and not a government agency. It is an open volunteer community that produces the standards the internet runs on: the protocols behind email, web pages, video calls. The work happens mostly online, on mailing lists, year-round; the three meetings a year are the weeks when the arguments get made face to face.
Who is in the room
Anyone. There is no membership and no national delegation. The people who show up are engineers from router vendors, cloud providers, universities, network operators, and open source projects, plus a steady number of independent participants. You register, you pay a registration fee (waivers exist), and you sit in the working group sessions. The IETF’s own material says it plainly: “Anyone can contribute to this work.”
Decisions in a working group are not votes. The unofficial motto is “rough consensus and running code” — the chair judges whether the room’s argument has settled, and a working implementation counts for more than a well-written objection. That sounds loose. In practice it has produced the internet’s core standards for forty years.
The drafts are public
Every working document is readable by anyone, without an account. They are called Internet-Drafts, and they live in the IETF’s Datatracker. The agentproto group already has around twenty of them — proposed protocols, architecture sketches, and use-case documents submitted by individual authors who want the group to adopt their approach. A few worth opening:
draft-agentic-ai-usecases-requirements-05 (October 27) describes what agents actually do across platforms. draft-manyfolks-agentproto-arch-00 (October 30) is a candidate reference architecture. draft-srivastava-agentproto-handoff-context-00 (October 26) proposes signed handoff context when one agent passes work to another. Each draft page shows its revision history, and the Datatracker’s diff tool shows exactly what changed between versions. When a draft matures and the group and the IESG approve it, it gets published as an RFC — the permanent, freely downloadable standards the rest of the internet builds against.
How to comment
This is the part most people do not know: the comment channel is an ordinary mailing list, and anyone can join it. For agentproto, you subscribe at ietf.org/mailman/listinfo/agentproto and email agentproto@ietf.org. The full archive is public, so you can read the arguments before saying anything. The group also has a Zulip chat channel and a GitHub organization where meeting materials and drafts are kept.
That is the whole access model. A company cannot outvote you, because there are no votes. A working group charter can exclude your topic, though — which brings me back to the hole.
The memory gap still stands
The agentproto charter defines the dialog context it will standardize so that it “does not include the content exchanged with AI models, such as conversation memory.” Memory is out of scope, by name. It matches the field: the two specifications that actually run in production (MCP and A2A) exclude memory too. In the registry of agent-memory standards I run at worldagrees.com/registry — fifteen specifications, drafts, and frameworks, tracked daily — not one has ratified memory scope.
So this week in San Francisco, the rules for agent conversations will get argued in a room anyone can enter, in documents anyone can read, through a list anyone can email. And the thing those conversations leave behind — what an agent remembers, who can read it, whether it can be proven untampered, whether it can be erased — is being decided by nobody in public at all. The process is open. The scope is not. If the memory question ever gets a standards home, it will start the same way this group did: someone subscribing to a list and writing the first draft. It is a smaller barrier than it looks.
The book’s last chapters argue that disagreement needs somewhere to go — a channel where “I do not agree with what this system has learned about me” actually lands (chapters 10 and 11). The IETF is one of the few places on the internet where that machinery is built in public. Watch this one.