A reading page, written at Martin Pompéry’s request (2026-10-26) to elaborate one section of Li et al., Agentic Communication Networks (2026). Its argument is that §5.4 of that paper restates, in a domain where SINE has no voice, an argument SINE has five years of evidence on from a domain where it does. Quotations are from the source; everything else is ⟨Inference⟩.
The proposal
§5.4. An Agent Exchange Gateway sits between heterogeneous agent protocols — MCP, A2A, ACP, and domain-specific ones — and translates each into a Protocol-Independent Intermediate Representation (PIIR) that “uniformly represents message metadata, task semantics, capabilities, context, interaction state, and security policies”.
The pipeline: protocol parsing → PIIR generation → target-protocol reconstruction → runtime validation. “Protocol knowledge graphs, semantic mappings, and template engines preserve task intent and interaction state during conversion.”
And the payoff claim, verbatim:
“A plug-in architecture allows a new protocol to be integrated by defining only its bidirectional mapping to the PIIR, reducing the complexity of supporting n protocols from O(n²) pairwise mappings to O(n).”
Candidate techniques named: abstract syntax trees, protocol state machines, semantic ontologies, protocol knowledge graphs, embedding-based alignment, LLM-assisted parsing, schema mapping, template engines, protocol plug-ins, runtime consistency verification.
Why this is familiar
⟨Inference.⟩ This is the hub-versus-point-to-point argument, which is the standard argument for a common data model in any n-party exchange problem. PACT is the hub in product carbon footprints: one wire format and data model, conformance against it, rather than bilateral integrations between every pair of trading partners. iLEAP describes its own contribution in exactly those terms — “the plumbing — how a shipper pulls emissions data out of a carrier’s system without a bilateral integration project.”
The paper asserts the O(n) result. SINE has run the experiment. Three things the clean form of the claim has to carry, each with a case behind it.
(a) O(n) holds only while the intermediate representation is semantically complete
The moment a protocol carries meaning the PIIR cannot express, implementers build bilateral profiles — and the n² mappings return, hidden inside “extensions”.
The paper knows the objection and parks it. §4.4, one subordinate clause: interoperability work proceeds “while recognizing that complete interoperability may not always be achievable because of capability mismatches or ambiguous and non-equivalent representations”. It is never held against the O(n) sentence a page later.
⟨Inference: the worked case is in this wiki. iLEAP rode inside PACT as a data-model extension until v1.1.0 — the extension mechanism being exactly the escape hatch the hub model needs in order to survive incomplete semantics. v1.1.0 makes a standalone iLEAP API mandatory for conformance and schedules the PACT path for removal in v2.0.0.
The honest statement of the case matters here: the stated cause of the split is PACT’s own release management — the PACT v2 specification family was sunset on 1 April 2026 — not a semantic failure. What the case does show is subtler and more useful. What survived the separation was a mapping table and an authentication flow. What did not survive, and was never claimed, is semantic interoperability: the iLEAP editors “strive to assert” it “with future releases”, and §8 of the specification states in bold that conformance “ensures interoperability and syntactic correctness” only.
So the lived finding is not “the hub collapsed”. It is: two standards, co-built by the same people, sharing a wire for years, never achieved the property the PIIR claim assumes as its premise. Syntactic translation was solved and semantic equivalence was held out of scope — and when the wire was withdrawn for unrelated reasons, there was no semantic layer left holding them together. A PIIR for agent protocols is being asked to do, by construction and across a larger and faster-moving protocol set, what the carbon-data case deferred.⟩
(b) The saving is bought with institutional centralisation
⟨Inference.⟩ Someone owns the PIIR, versions it, and decides what is expressible in it. O(n) is purchased by creating a single point of control over which meanings can exist in the ecosystem at all. The paper presents the plug-in architecture as a purely technical property; it is a governance architecture with a complexity result attached.
The paper half-notices this one section earlier. §5.3 observes that capability ontologies, semantic embeddings, protocol mappings, task models and policy vocabularies “require explicit version and compatibility identifiers, freshness and locality constraints, distribution policy, and resource accounting so that an agentic service can be rebound or migrated without silently changing the meaning of its intent, capabilities, or task state” — citing KB-MANO [41], which treats shared knowledge-base lifecycle as a joint communication–computing decision. That is a description of versioned, governed, distributed semantics. It is not connected to §5.4.
⟨Inference: this is where SINE’s actual competence sits. In carbon data the schema was never the hard part; the governance of the schema was — who convenes, who decides what is in scope, how versions sunset, who bears the cost of a breaking change. See Data Commons and Ostrom’s design principles. The transferable claim is that the intermediate representation is the product, and its governance is the business model — and that an architecture which does not name its convener has not finished its design.⟩
(c) The translation is proposed as probabilistic, across a boundary where that is not safe
⟨Inference.⟩ Embedding-based alignment and LLM-assisted parsing are mappings without guarantees. The consequence depends entirely on what is being mapped:
- A lossy translation of a QoS parameter degrades a stream. Recoverable, observable, bounded.
- A lossy translation of an authority or a constraint changes who is permitted to do what. Not observable at the point of failure, and the failure is silent.
The PIIR is specified to carry security policies alongside message metadata. So the probabilistic path runs straight through the authorisation boundary — which the companion page Agent identity and delegated authority shows is already the weakest joint in the architecture.
The paper holds the pieces of this objection and never assembles them: §5.3’s warning against “silently changing the meaning”; §4.4’s admission of “ambiguous and non-equivalent representations”; and its own MCP/IoRT citation [42] in §5.4, which reports that semantic interpretation and tool invocation “account for much of the workflow latency” and concludes that MCP “complements rather than replaces deterministic middleware for strict real-time or safety-critical control”. The paper’s own recommendation there — “preserve typed capability schemas, authority, task context, and latency class across protocols, while keeping time-critical execution on a bounded and verifiable control path” — is the right instinct and contradicts the LLM-assisted translation path proposed in the same section.
What SINE could say that nobody is saying
⟨Inference. A position, not a finding — nothing below is sourced to an output SINE has published.⟩
Three transferable claims from the carbon-data experience, none of which currently appears in the agent-protocol conversation:
- The common model wins where the semantics are thin and the incentive to defect is low. It is a coordination device, and it works on coordination problems. See coordination vs. bargaining failure.
- Extension mechanisms are where interoperability is decided. They are presented as flexibility and function as the exit. An architecture’s real n² count is the count of its profiles, not of its protocols.
- The intermediate representation is a governance object. Version policy, sunset policy and convening authority determine whether O(n) survives contact with a second implementer.
⟨Inference: this is also the most concrete bridge between SINE’s two centres of gravity — the sustainability-data standards work and CCTI — which the CCTI page records as thin. The bridge is not a shared technology; it is a transferable finding about how shared representations are governed.⟩
Related
- Li et al., Agentic Communication Networks (2026) — the source
- Agent identity and delegated authority — the other half of the same source
- PACT Pathfinder Network · iLEAP · iLEAP Technical Specifications (2026)
- Interoperability profile — “an architecture’s real n² count is the count of its profiles”, generalised ⟨2026-11-06⟩
- Data Commons · Ostrom’s commons design principles
- Coordination failure vs. bargaining failure
⚠ ⟨Flagged by the 2026-11-02 lint — this page has only heard one side.
Every claim above is drawn from a single source, Li et al.,
a research proposal. Meanwhile sources/standards/2026-05-22-gbz-185.1-2026-agent-interconnection-general-architecture/
— GB/Z 185.1—2026, Artificial intelligence — Agent interconnection —
Part 1: General architecture, a Chinese national guiding technical document
— has been in sources/ since 2026-10-12, two weeks before this page was
written, and is un-ingested and unregistered (open task 11).
A published national standard is a different kind of evidence from a paper:
it is a convening authority that already exists, which is precisely the
variable section (b) argues the saving is bought with. Read it before
treating the three findings above as settled.⟩