Three agent specifications were promised for October 2026. As of the tenth of that month, none had shipped. The market has settled on a comfortable interpretation: this is a delay, a slipped deadline, a schedule risk. That interpretation is wrong, and it is wrong in a way that matters to anyone holding capital in the infrastructure layer.
The three deliverables are not equivalent objects. One is a protocol working draft. One is a product general-availability release. One is a registry specification. Wrapping them together as "three unshipped specs" does rhetorical work that the underlying facts do not support. It inflates the delivery gap and, more damagingly, it hides where the stack is actually failing. The stack is not failing at the standard layer. It is failing at the identity and trust layers, where no convergent design exists and no mandatory conformance test forces one.
I have seen this exact misreading before. In late 2017 I declined three high-profile token sales while the market priced whitepapers as products, and I spent four hundred hours auditing a DeFi prototype instead. The market was pricing announcements. I was pricing the missing implementation. The gap between those two numbers is where capital is lost. Architecture reveals the true intent โ and the architecture here says the announcement is not the product.
To understand why the "delay" framing is misleading, the stack has to be laid out as a stack. Agent infrastructure in 2026 organizes into roughly seven functional layers. At the top sits the application and interface layer, where consumer-facing agents live. Beneath it sits orchestration and memory. Below that, commerce and payment. Below that, trust and authorization. Below that, identity and authentication. Below that, discovery and registry. And at the base, transport and distribution.
The three "unshipped" items do not live on the same floor. The Gemini Agent is a product release spanning the application and orchestration layers โ a general-availability product, not a specification. The A2A Registry is a discovery-layer artifact. The Poppy working draft, formally the Poppy Agent Protocol, spans the commerce and identity layers. Calling all three "agent specs" is a category error, and category errors in infrastructure analysis are not cosmetic. They misdirect the capital that follows them. When a fund allocates to "the agent standardization theme," it is allocating to three different floors with three different risk profiles, and it does not know it.
The distribution layer, meanwhile, is already solved. The Docker-and-OCI analogy that surfaced in the discussion is the most technically honest observation in the entire debate: OCI standardized how containers are distributed and transported, and said nothing about agent-level trust semantics, permission manifests, or sub-agent guarantees. That gap is real and it is load-bearing. MCP solved tool-call transport. A2A solved agent-to-agent message transport. Neither defines the question that actually governs whether an autonomous agent may spend your money: is this agent authorized to represent me in this transaction, under what constraints, and who attests to that authority?
That question sits at the identity and trust layers. It is unanswered. That is the vacuum the vendor extensions are rushing to fill. The fragmentation story is being told as a standards failure. It is better understood as a land grab at the identity layer, and the land grab is happening precisely because no one has published a convergent identity design.
Start with the Poppy draft's identity foundation. One of its five building blocks is OAuth-based login. The discussion treats OAuth as a solved module and moves on. That is the single most consequential technical blind spot in the whole debate. OAuth 2.0 was designed for delegated authorization of human users. Its known deficiencies in an agent context are not minor: it lacks fine-grained, composable permission credentials; its tokens are routinely over-scoped; its revocation granularity is coarse; and it cannot natively express constraints of the form "agent A acts for user U, in context S, up to a ceiling of X." It also has no native mechanism for agent-to-agent delegation chains. OAuth as the identity substrate for autonomous agents is not a foundation. It is an assumption wearing a foundation's clothes.
I built this argument into my own 2026 research on the AI-crypto convergence, specifically when I analyzed the verifiable-compute mechanism of a zero-knowledge AI protocol. The finding that mattered was not the proof system. It was that without a cryptographic attestation of what an agent is authorized to do, machine-to-machine settlement has no trust anchor at all. The proof that a computation ran correctly is a different object from the proof that the agent running it was entitled to run it. The first is being solved. The second is being ignored. The ledger remembers what the market forgets โ and what the market is forgetting is that identity, not compute, is the binding constraint.
There is a second-order problem the discussion does not reach. Even if the identity layer were specified tomorrow, the settlement layer beneath it has no way to consume an agent's authorization state. On-chain, a smart contract can verify a signature and a balance. It cannot, by itself, verify that the signing agent was operating within the bounds its principal set. Bridging that requires the authorization state to be committed to a verifiable form โ a credential that a contract or a verifier can check without trusting the agent's own report. The zero-knowledge tooling to do this exists in fragments. The standard that would make it composable does not. So the agent stack is not merely missing an identity spec; it is missing the bridge between identity and settlement, and that bridge is the layer where crypto and AI actually meet.
Now the discovery layer, where the path collision lives. Poppy proposes a well-known URI at /.well-known/poppy.json. The well-known URI space is scarce and mutually exclusive by design. If Poppy occupies one path, A2A uses another, and the MCP ecosystem uses its own, a single domain may be forced to maintain multiple overlapping discovery files with partially duplicated semantics. This is not a publication problem that a v0.1 release would fix. It is a coordination problem that persists even if every draft ships on time. The discussion does not raise it.
The pattern is familiar from crypto. Mapping the invisible currents of liquidity taught the same lesson in 2020: the visible layer โ the exchange, the front end โ was never the risk. The risk was in the plumbing between layers, where the semantics of one system's output did not match the semantics of the next system's input. Agent discovery has the same disease. The registries will publish. The semantics will not line up.
Here is where I want to apply the discipline I use on-chain. On October 6, the Poppy draft announced six launch partners. On October 9 โ three days later โ it added thirty-five design partners. The market reads "forty-one partners" as proof of momentum. Read it as a data series instead. A partner count that grows from six to forty-one in seventy-two hours is not the signature of technical consensus forming. It is the signature of a coalition list being assembled, with the exclusivity and the optics that coalition lists carry.
Signal extraction from the noise floor is the whole job here. Partner counts are a noise-floor signal. They correlate with press coverage and with nothing else that predicts whether an interface will actually be adopted. In 2017 I watched token sales with fifty "strategic partners" fail on contract logic. The partners were never the product. The partners were the marketing of the product. The same discount applies now, and the discussion's decision to cite the partner count as evidence of the draft's weight is a misreading of the signal's nature, not merely of its magnitude.
The memory taxonomy shows the same slippage between engineering and packaging. The Gemini Agent's four memory classes โ session, semantic, procedural, and episodic โ map almost exactly onto cognitive psychology: Tulving's semantic and episodic memory distinction, plus procedural memory. Importing a forty-year-old psychology taxonomy into a product narrative is not an engineering disclosure. It is a branding decision. The actual engineering problems in agent memory are storage tiering, cross-tenant isolation, memory-poisoning defenses, and the right to deletion under GDPR. None of those appear in the taxonomy. None of them appear in the discussion. A four-word list is doing the work that a storage architecture should be doing.
And the claim that multi-agent orchestration can run "for hours to days" should be read as a capability claim, not a maturity proof. Long-running agents have four unsolved engineering problems: state persistence, failure recovery, cost control, and observability. "Hours to days" does not solve them. It amplifies them. Every additional hour of autonomous operation is another hour of hallucination accumulation, cost drift, and permission drift. Certainty is a liability in this domain, and a duration claim is the most seductive form of false certainty in agent infrastructure. The risk surface of a two-hour agent run is not twice the risk surface of a one-hour run. It is nonlinear, because the failure modes compound, and the market is pricing them as if they were additive.
Strip away the packaging and one technical fact survives cleanly: the distribution layer is solved, and the trust and identity layers are not. The discussion gets this right, and it deserves credit for it. What it does not say is why the vacuum persists even though the need is obvious.
The answer is that there is no mandatory conformance suite. Standards without conformance testing are suggestions. A specification that describes an identity layer but cannot test whether an implementation honors it is a document, not a protocol. In cryptography, an algorithm is only as strong as its test vectors and its reference implementation. Agent identity has neither a published reference implementation nor a cross-protocol test suite that would force two vendors' agents to actually interoperate under a common trust model. That absence is the reason vendor extensions are proliferating: without a test that says "this extension is non-conformant," every vendor extension is technically admissible, and the stack fragments by default rather than by decision.
This is the one question the discussion raises and then drops: who issues the trust anchor for the Trust-Authorization Overlay? Is it the platform vendor, the domain holder self-signing, or a third-party certificate authority? That single answer determines whether the agent identity layer becomes an open utility or a single-vendor chokepoint. The discussion flags the question and moves on. The question is the whole ballgame. Patterns repeat, but the participants change โ the fight over who signs the certificate is the same fight that played out over who runs the certificate authority on the web, and it is being re-run with higher stakes, because the certificate here authorizes spending, not merely browsing.
The consensus reading is that fragmentation is a symptom of delay โ that if the specs had shipped on time, vendors would not be filling the vacuum. Reverse it. The vacuum is not an accident that the delay created. The vacuum is the objective. A vendor that ships its own extension into an unstandardized layer sets the default that the eventual standard must accommodate, and the vendor that sets the default shapes the standard. Slow standardization does not hurt the strongest vendor in the room. It helps it. The consensus is often the contrarian trap, and the consensus here โ "they missed the deadline, how embarrassing" โ is precisely the framing that a dominant vendor would prefer the market to hold.
Crypto already ran this experiment, and the results are in the ledger. Exchange "proof of reserves" is the canonical case: an exercise that proves part of the liabilities and none of the solvency, dressed as transparency, and accepted by a market that wanted the announcement more than the audit. Layer 2 sequencers are the second case: marketed as decentralized for two years while running as single centralized nodes, because the decentralization was a roadmap slide and the centralization was the shipped product. In both cases, the standard was late, the vendor default filled the gap, and the default hardened into the market structure. Agent identity is walking the same path. The vendor extension is not filling a temporary vacuum. It is pouring concrete.
The question to hold into the next quarter is not "when will the specs ship." It is "who signs the trust anchor, and who publishes the conformance suite that makes it binding." If the answer to both is a single platform vendor, then the agent stack will not have standardized. It will have consolidated, and the consolidation will have been financed by a market that mistook a slipped deadline for a temporary inconvenience. Survival is a function of position sizing โ and the position to size is not the spec that ships late. It is the identity layer that never converges.