Three specifications were promised for October 2026. On October 10, none had shipped. The vendor extensions filling the vacuum are not a stopgap. They are a fork.
That is the surface read. The forensic read is different.
I pulled the timeline on the one artifact that did move: a protocol working draft called Poppy. It announced six launch partners on October 6. Three days later, on October 9, it appended thirty-five design partners. A forty-one-name coalition assembled in seventy-two hours is not a technical consensus. It is a recruitment list with a logo wall. I have audited enough ICO partner pages to recognize the pattern: partner count is a marketing metric, not a readiness metric.
The bytecode lies; the transaction log does not. And the log here records a gap between announcement cadence and shipping cadence. That gap is the story.
Let me set the methodology before the claims. The source material for this analysis is an unverified industry report dated October 10, 2026. It carries no byline. Eight of its factual anchors have no traceable origin. Named individuals, events, and registries β kthota-g, Carles Arnal, Gemini at Work 2026, the AI Catalog, the ARD β cannot be cross-checked against any public record. One factual error is already visible: the report calls Thomas Kurian "Google CEO." He is Google Cloud CEO. Sundar Pichai runs Alphabet. A single title error does not invalidate a document, but it downgrades the document's verification density by one tier.
So I analyze logic, not truth. I treat every claim as a claim. This is the same discipline I apply to unaudited token contracts: I do not price the narrative; I trace the execution path. The marketing wrapper is irrelevant; the call graph is not.
The subject is the agent protocol stack β the layered infrastructure that lets autonomous software agents discover each other, authenticate, delegate authority, and transact. Think of it as the TCP/IP of machine commerce. The layers, bottom-up:
L1 transport: HTTP, OCI, A2A transport. Solved. L2 discovery: registries, catalogs, /.well-known/ files. Contested. L3 identity: OAuth-based login. Fragile. L4 trust and authorization: permission manifests, delegation chains. Unresolved. L5 commerce: payment and settlement modules. Draft stage. L6 orchestration and memory: multi-agent scheduling. Product capability. L7 application: user-facing agents. Shipping.
The report collapses all seven into a single headline β "three unshipped specs." That is the category error I will dismantle.
Here is the central finding. The three deliverables framed as one class of failure are three different classes of artifact.
One is a protocol working draft β a standards document. One is a product general-availability release β a shipping binary. One is a registry specification β a naming and discovery layer. A draft, a product, and a registry are not peers. They live at L5, L7, and L2. Presenting them as "three specs" inflates the rhetorical weight and, more importantly, obscures where the actual bottleneck sits.
The bottleneck is not the spec layer. It is the trust layer. That distinction determines whether the ecosystem fragments or consolidates.
The most technically honest sentence in the source is its Docker analogy: OCI standardized distribution and transport but said nothing about agent-level trust semantics, permission manifests, or sub-agent guarantees. That judgment is correct, and it maps cleanly onto what I see in adjacent ecosystems. MCP standardized tool invocation. A2A standardized inter-agent messaging. Neither defines the question that matters: is this agent authorized to execute this transaction on my behalf? That is precisely the layer the Poppy draft claims to fill. And it is the layer with no convergence.
Now the blind spot the source skips entirely: it treats OAuth as a solved module. OAuth 2.0 was designed for human-delegated authorization. Its failure modes are documented and severe when you push it into machine-to-machine delegation:
It lacks fine-grained, composable permission credentials. You need RAR, GNAP, or UCAN extensions to express them. Tokens are chronically over-scoped, and revocation is coarse β all or nothing. It cannot natively express a constraint like "Agent A, acting for user U, in context S, capped at amount X." It has no native support for agent-to-agent delegation chains.
Wrapping OAuth in an agent wrapper does not fix the substrate. It inherits every flaw. The source treats the weakest assumption in the entire stack as a settled component. Trust the hash, verify the execution path β and the execution path here runs straight through an authorization primitive that was never built for autonomous machines.
Second blind spot: the discovery layer. The draft reserves /.well-known/poppy.json. Meanwhile A2A occupies /agent.json, the MCP ecosystem maintains its own path, and the broader agent community has agents.json. The .well-known URI space is a scarce, mutually exclusive resource. Reserve one, and you may be forced to host multiple overlapping discovery files under a single domain. This is not a "spec not shipped" problem. It is a coordination problem that persists even after the spec ships. The source never raises it.
Third: the memory taxonomy. The product narrative cites four memory types β conversational, semantic, procedural, episodic. That is a direct lift from cognitive psychology: Tulving's semantic/episodic distinction plus procedural memory. It is a packaging taxonomy, not an engineering disclosure. The real engineering questions β storage tiering, cross-tenant isolation, memory-poisoning defense, GDPR deletion rights β go unmentioned by both the vendor and the source. When a system describes itself in the vocabulary of the human mind, the audit surface has moved from code to copy.
Fourth: the long-running agent claim. The source notes that multi-agent orchestration can "run for hours to days." That is a capability statement, not a maturity proof. Long-duration agents raise exactly the risks I model in stress tests: state persistence, failure recovery, cost containment, observability. Hours-to-days runtime amplifies hallucination accumulation, budget overrun, and permission drift. Pressure tests expose what calm markets hide. A demo that runs for two days is not evidence of a system that survives two days of adversarial input.
Correlation is not causation, and partnership is not readiness. I have watched a thirty-partner consortium dissolve inside a quarter once its first independent audit failed.
The dominant framing β "a vacuum, filling with fragmentation" β depends on reading silence as absence. But silence in the logs speaks louder than tweets. The source never mentions that MCP is already a de facto standard at the tool layer, or that the A2A protocol itself β distinct from its contested registry β has real adoption. By scoping the story to the unshipped artifacts, it manufactures a vacuum that the installed base contradicts.
Data does not dream; it only records. And what the record shows is not a stalled field. It shows a field where the easy layers β transport, tooling β have consolidated, and the hard layers β identity, trust, delegation β have not. That is normal protocol maturation, not collapse. TCP/IP took a decade to settle its security layer. The agent stack is three years in.
The fragmentation is real. But it is fragmentation at the trust layer, where it is expensive, not at the spec layer, where it is cosmetic. Volatility is noise; structural flaws are signal. The structural flaw is that no one has defined who issues the trust anchor for a cross-vendor agent authorization. Is it the vendor? The domain holder self-signing? A third-party CA? Until that question is answered, every extension filling the vacuum deepens the fracture.
Watch the conformance suite, not the press release. The signal next week is not another partner count. It is whether a v0.1 draft ships with an open-source reference implementation, a named license, and a mandatory consistency test. Reproducibility is the only currency of truth. An agent standard you cannot run against a test vector is a PDF, not a protocol. When the trust anchor gets a signer, the stack converges. Until then, the vacuum is being filled by fragmentation β and the fragments do not compose.