Summary

  • The posted Agentproto minutes call the eight show-of-hands results a nonbinding sense of the room, with consensus to be confirmed on the mailing list. They do not record a charter approval or an established Working Group.
  • The room rejected the proposition that the initial scope was correct by 38–124–40 and split 92–58–49 on the initial deliverables, while answering 154–51–5 in favour of forming a WG “with this charter.” Those predicates cannot be collapsed into one approval score.
  • The totals vary by question and the public table contains no respondent-level cross-tab. It cannot show how many people who opposed the scope also supported formation, or whether participants gave the phrase “with this charter” the same meaning.
  • The charter shown at the session can be frozen at repository commit a9405f3d5d0a4e8c67342d7945137c8230b9cd05. Later emails and an open, unmerged scope-reduction pull request show that important boundaries remained under active discussion.
  • Agentproto remained in Datatracker state BOF with no charter at the research cutoff. Room evidence, list confirmation, Area Director judgment, IAB advice and IESG approval are different acts.
  • A public predicate-reconciliation matrix would preserve those distinctions without publishing named ballots: exact question, immutable text, response denominator, objections, required edits, list disposition and final authority state.

Two answers, not one mandate

The neat story would be that the Agentproto BoF endorsed a new Working Group. It has a number that looks decisive: 154 yes against 51 no. The opposite neat story would be that the BoF rejected its own charter, because 124 people said the initial scope was not correct while only 38 said it was. Both stories remove the grammar from the evidence.

The posted minutes preserve a more complicated record. They describe a WG-forming BoF held on 23 July 2026, label their own status “DRAFT for chair review,” and say the results are a sense of the room taken by show of hands through Meetecho. They expressly call the exercise nonbinding and say consensus is to be confirmed on the mailing list. Attendance figures were not yet available. The minutes also disclose that an AI agent harness helped compile the draft from human notes, the session chat and slides, with final correction remaining the chairs’ responsibility.

That provenance does not make the minutes useless. It tells the reader what they are: an important provisional institutional record with clear limits. It should be quoted with its status, not upgraded silently into a final transcript.

The published agenda supports the session’s intended sequence. Technical motivations and proposed components came first; open-mic discussion followed; consensus calls were scheduled near the end. The IETF’s pre-meeting BoF guide framed working-group-forming BoFs as tests of whether there is enough consensus, energy and sufficiently tight scope to charter work. That is a diagnostic function. A diagnosis may expose several different conditions at once.

The complete table does exactly that. The problem was judged well understood, solvable and useful by 124–72–14. The need for interoperability received 155–30–8. The IETF as the right venue received 158–30–14. Initial scope received 38–124–40. Initial deliverables received 92–58–49. Forming a WG “with this charter” received 154–51–5. Willingness to write or review received 137–17–40. Willingness to implement received 42–28–40.

The minutes’ own synthesis is careful. They say the room showed strong support for the value of the problem, the need for interoperability and the IETF as the venue, as well as support for forming a Working Group. They also say the initial scope as written was not supported and views on the deliverables were split. At wrap-up, Laurent Ciavaglia cautioned that asking whether scope and deliverables were “correct” should not be confused with asking whether they were “complete.” The chairs would take the results and the charter discussion to the list.

That is not a failed record. It is a conditional one.

Why the results can coexist

A participant can believe that a durable IETF venue is needed while opposing part of the proposed boundary. A participant may support formation on the expectation that the charter will be repaired. Another may read “with this charter” as shorthand for a recognisable charter family rather than approval of every clause. A third may think the scope is incomplete, not wrong, and still answer no to “correct.” A fourth may support the work in principle but oppose the particular deliverables. None of those positions is irrational.

The record, however, does not tell us which explanation applies to how many respondents. Question totals range from 110 to 210. Questions 4 and 6 each have different response structures even though both total 202 and 210 respectively. The published table contains no per-person joint distribution. It therefore cannot support the sentence, “The same 124 people rejected the scope and then voted to form the group.” Nor can it show that none of them did.

This matters because aggregate marginals do not reveal the relationship between answers. If 124 people said no to scope and 154 said yes to formation, the overlap could be substantial, small or somewhere between. Publishing a respondent-by-respondent ballot would be alien to the purpose of rough-consensus judgment and could chill participation. But the chairs can still publish an aggregate reconciliation: for example, whether the dominant reason for a no on Question 4 was excess breadth, missing work, ambiguous terms or a particular deliverable; and whether support on Question 6 was understood as conditional on named edits.

The distinction is not statistical fussiness. It is what stops a headline number from becoming a substitute for the institutional judgment the session was designed to inform.

“With this charter” is the difficult phrase

The strongest defence of the result is that BoF questions are deliberately piecewise. RFC 5434 recommends questions that locate agreement and expose where further work is needed. On that reading, Questions 4 and 5 diagnose components while Question 6 asks about the project as a whole. It is entirely possible for the whole to be worth creating even when components require repair.

But RFC 5434’s example is more exact than “Should this topic have a home?” It asks whether there is support to form a WG with the following charter, and explains that this means the charter itself is ready as shown by community support. The Agentproto wording used the same connective tissue: “with this charter.” That makes Question 6 an integrated proposition, not a free-standing poll on whether agent communication is interesting.

The answer is not to declare the result invalid. It is to preserve the condition under which the chairs interpret it. If “yes” meant “form a WG after repairing the rejected scope,” the record should name the repair. If “correct” was heard as “complete,” the record should identify which omissions mattered. If the final charter differs materially from the text shown, the later synthesis should say which BoF objection caused each change and whether list discussion resolved it.

If the IESG eventually approves a narrower or differently structured text, the public chain should not imply that 154 people had already endorsed that later text in Vienna.

Without that chain, one can quote Question 6 as blanket authorization. Someone else can quote Question 4 as a veto. Both quotations would be numerically accurate and institutionally incomplete.

Freeze the text before interpreting the room

The minutes link a proposed charter and say current charter text was displayed during the session. The official ART-area repository supplies an unusually useful anchor: the charter at immutable commit a9405f3…. That commit created the file on 21 July, two days before the BoF. At the research cutoff, the repository history returned no later commit to that path.

The frozen text is broad enough that “scope” was not an abstract word. It proposed work on agent-to-agent and agent-to-tool communications; specified protocol mechanisms for human-agent sessions; described identity, authorization, confidentiality, integrity and auditability considerations; and named a standards-track session protocol plus an informational framework. The session protocol would cover context handling, long- and short-lived sessions, recovery, multiple data timing classes and point-to-multipoint communication. The framework would explain how agent building blocks compose with work elsewhere.

Freezing that file does not prove that every respondent read every line or gave each term the same meaning. It does something more modest and essential: it prevents the object of the poll from changing after the fact. “The charter” can point to a blob, not a moving branch.

This is where versioning becomes governance rather than repository hygiene. If a later charter removes session lifecycle, narrows point-to-multipoint work or reframes the deliverables, then it is not textually identical to the object placed before the room. That may be the right outcome. The change simply needs a traceable reason and an authorized later decision.

The post-BoF record confirms that the boundary was still moving

The public discussion after Vienna makes it unsafe to treat 154 as approval of a settled scope. On 12 August, Suresh Krishnan’s list message, “Sussing out points of disagreement on the charter”, said the comments had prompted two ground-up rewrites. It identified six live questions: substrate versus semantics, session lifecycle, correlation versus proof of binding, point-to-multipoint versus bilateral communication, the number and status of deliverables, and how much terminology the charter should define.

Those are not copy edits. They describe the nature of the output, the lifecycle being standardized, the topology, the strength of proof and the work programme.

The official repository’s pull request 77, opened after the BoF against the meeting-era commit, proposed a “minimum viable primitive” reduction. Its description would move session lifecycle out of scope and turn the session-protocol deliverable into an interaction reference and binding envelope. At the cutoff it remained open and unmerged. The description also says its formulation relied on AI-assisted chat-log analysis. That makes it a public proposal with useful provenance, not an approved replacement and not a chair determination.

A later list proposal on evidence requirements offers a smaller example of the same issue. Nancy Sahu suggested clarifying what operational evidence would establish and what it could not establish, and tying the charter’s auditability consideration to retention of an evidence clause in a deliverable. The point here is not whether that wording should win. It is that even a single advertised property can depend on whether a corresponding deliverable survives revision.

The IETF 126 retrospective published on 1 September says Agentproto tested consensus for chartering and that discussion continues on the mailing list. The current Datatracker group page supplies the decisive null: state BOF, Charter (None), “Not chartered yet.” A null is not failure. It is the current authority state.

A room informs authority; it does not finish it

RFC 2418 / BCP 25 describes the formal chain. A prospective charter is negotiated with the relevant Area Director, then reviewed with IAB advice and finally approved, changed or declined by the IESG. If approved, the Working Group and its charter are recorded and announced. The same document also warns that 51% is not rough consensus and discusses mailing-list verification of face-to-face conclusions.

This design deliberately resists both pure vote-counting and private fiat. Counts reveal support, opposition, participation energy and where questions were misunderstood. Chairs interpret the substance of objections. The list allows people absent from the room to respond and lets positions be tested in writing. Area Directors and the IESG carry defined institutional responsibilities. None of these layers makes the others decorative.

Heng Lu’s critique of attendance becoming mandate is useful here only as a discipline of language. Participation can be evidence, expertise, warning and objection. It should not be represented as authority over a decision that still belongs to another actor. Applied narrowly to Agentproto, that means the room’s result deserves preservation precisely because it is evidence—not because it can be inflated into a charter.

The same restraint protects the room. If the IESG later chooses a different boundary, it should not rewrite the BoF as having endorsed that text. If the project is not chartered, that should not erase the strong support for interoperability work or willingness to contribute. Each claim keeps its own source and clock.

The predicate-reconciliation matrix

The missing public object is not another meeting summary. It is a compact matrix that lets a reader reconstruct how one proposition became the next.

For each consensus question, the first column should carry a stable identifier and the exact wording. The second should say what institutional decision the question was meant to inform. The third should link the immutable charter commit and blob displayed. The fourth should record session time, collection method, yes, no, no-opinion and total responses, together with the explicit nonbinding status.

The next columns should record interpretation rather than pretend it was arithmetic: terms identified as ambiguous; principal objection themes; whether “correct” was distinguished from “complete”; the clauses that required revision; and whether support was understood as conditional. Named individual ballots are unnecessary. A small set of chair-attributed categories and a privacy-safe aggregate cross-tab, if it can be produced without reconstructing identities, would be enough.

Then comes the authority chain: date and interval of mailing-list confirmation; links to the relevant threads; chair synthesis and objection disposition; responsible Area Director action; IAB advice; IESG decision; and the immutable final charter commit, or an explicit null while none exists. Corrections and superseding interpretations should append rather than overwrite the earlier state.

For Agentproto, the matrix would currently show a genuine asymmetry. Problem value, venue and formation have positive room evidence. Initial scope lacks support in the posted result. Deliverables are mixed. List refinement is active. The final charter and IESG approval fields are null. That is more informative than either “the BoF approved Agentproto” or “the BoF rejected Agentproto.”

Objections worth preserving

The first objection is that rough consensus is not a referendum. A matrix full of numbers could tempt readers to calculate a winning percentage and ignore the quality of objections. The answer is to label counts as evidence and give chair reasoning, unresolved objections and authority state equal prominence. A percentage should never be the status field.

The second is that chartering is iterative. Requiring an immutable anchor might appear to freeze a living discussion. It does the opposite. A versioned chain allows change without pretending continuity of text. The old version remains the object of the old question; the new version becomes the object of a later review.

The third is privacy. Publishing named answer combinations could expose factions, intensify employer pressure and convert an engineering process into electoral analytics. The matrix does not need identities. It needs the exact proposition, totals, non-identifying objection categories and an accountable synthesis.

The fourth is workload. Chairs already have to read discussion, refine charters and report outcomes. A second elaborate dossier could slow useful work. The matrix should therefore be generated from artifacts already produced: minutes, charter commits, list threads, Datatracker events and IESG records. If the join requires manual historical reconstruction months later, that is evidence that the current record is already too fragmented.

The fifth objection is substantive: perhaps Question 6 was clear to everyone in the moment. Even so, the public record now contains Question 4 beside it and later documents major disagreements. A short explanation of the conditional synthesis would strengthen, not second-guess, the chairs’ judgment.

Evidence limits

This analysis does not determine whether the IETF should charter Agentproto. It does not choose between a session protocol, an interaction reference, a framework, a semantic layer or work in existing groups. It does not assess the quality of MCP, A2A, ACP, ANP or any implementation.

It does not establish who answered each question, how many unique people participated across all questions, whether anyone changed position, or whether a response was cast remotely or in the room. The totals cannot supply those facts. It does not infer bad faith, manipulation or procedural invalidity from the apparent tension.

The minutes remain draft for chair review. Their AI-assisted compilation is disclosed, but the checked sources do not show that any particular count is erroneous. Repository commit a9405f3… is the best fixed public anchor for the linked charter; it cannot prove subjective respondent understanding. Emails and pull requests show unresolved proposals, not outcomes. The Datatracker snapshot can change after the cutoff.

The finding is narrower: a public institutional record now supports different propositions with different numbers, while one integrated question refers to a charter whose scope component lacked support. That record becomes trustworthy when the predicates remain attached to text, objections and later authority. It becomes vulnerable when only the largest number survives.

Sources

  1. Heng Lu — The Multi-Stakeholder Mirage
  2. IETF 126 — Agentproto BoF minutes
  3. IETF 126 — Agentproto agenda, version 01
  4. IETF — IETF 126 Highlights
  5. IETF — Birds of a Feather at IETF 126
  6. IETF ART-area charters — Agentproto proposed charter at commit a9405f3…
  7. IETF Datatracker — Agentproto current group page
  8. RFC 5434 — Considerations for Having a Successful BOF Session
  9. RFC 2418 / BCP 25 — IETF Working Group Guidelines and Procedures
  10. Agentproto list — Sussing out points of disagreement on the charter
  11. Agentproto list — Proposed charter text: evidence requirements
  12. IETF ART-area charters — Pull request 77