On October 6, PAP published a list of six launch partners. Three days later, on October 9, that list had grown to forty-one โ thirty-five new "design partners" added in seventy-two hours. As of October 10, the draft itself โ poppy.json, the OAuth login module, the commerce block โ remains unshipped. So does the A2A Registry. So does the vendor-neutral trust overlay that was supposed to sit between them and hold the stack together.
I've watched this movie before. In 2017 I watched whitepapers move markets while the code sat in a branch that hadn't merged in months. In 2022 I watched an algorithmic stability narrative hold a forty-billion-dollar market cap four days before the oracle broke. The tell is always the same: when a project's partner count grows faster than its commit count, you're not watching standardization. You're watching marketing with a technical glossary.
Here's the number that matters: forty-one partners, zero shipped specs, and no conformance suite anywhere in sight. That's the actual state of the agent stack on October 10, 2026. Everything else is noise.
Let me set the table, because most of the coverage on this has the geometry wrong.
The story being sold is simple: three agent specs promised for October remain unshipped, and the vendor extensions rushing in to fill the vacuum are fragmenting the stack. Clean narrative. Three deliverables, one deadline, one failure.
The problem is that the three deliverables aren't the same kind of thing. Not even close.
PAP is a protocol work draft โ a specification. Gemini Agent is a product that already went GA. The A2A Registry is a registry spec layered on top of a transport protocol that's already shipping. Stacking a product launch next to two specs and calling all three "Agent Specs" is a category error. It inflates the failure, and worse, it hides where the failure actually lives.
I run a copy-trading community โ a thousand retail traders mirroring a rules-based strategy I built after losing $400,000 on Terra. I don't trade narratives. I trade structure. And structure requires knowing which layer of a stack is load-bearing before you decide where the fault line runs.
So let's do the layer analysis. If you don't know which layer you're standing on, you're not trading. You're gambling with extra steps.
The stack, as it actually decomposes, runs seven layers.
L7 is application and UI โ Gemini Agent. A product. It shipped. Calling a GA'd product an "unshipped spec" is nonsense.
L6 is orchestration and memory โ the four memory types, the multi-agent runtime. Product capability.
L5 is commerce and payment โ the PAP commerce module. Protocol draft.
L4 is trust and authorization โ the A2A Trust-Authorization Overlay. Spec component.
L3 is identity and authentication โ PAP's OAuth login. Protocol draft.
L2 is discovery and registry โ A2A Registry, AI Catalog, ARD, poppy.json. Spec plus multiple implementations.
L1 is transport and distribution โ OCI, HTTP, A2A transport. Solved.
Now watch what happens when you sort the "three unshipped specs" into this stack. The commerce draft lives at L5. The identity draft lives at L3. The registry lives at L2. And the product that supposedly shipped late lives at L7.
You cannot compare an L7 product GA to an L3 identity spec and call the delta a single "delivery gap." That's like declaring your car broken because the engine shipped but the license plate didn't. The headline number โ three โ is a rhetorical construct, not a technical one.
Here's the part the coverage gets right, and it's the most load-bearing sentence in the whole narrative: OCI standardized distribution and transport, but said nothing about agent-level trust semantics, permission manifests, or sub-agent guarantees.
That's accurate. MCP solved tool-call transport. A2A solved agent-to-agent messaging. Neither one defines whether this agent has the right to execute this trade on your behalf. That's the vacuum. That's what PAP claims to fill. And that vacuum is real โ I'll give the narrative that.
The question is whether the thing filling it actually works. It doesn't. Not yet. Let me show you why, layer by layer.
L3 is the weakest link, and almost nobody is talking about it. One of PAP's five building blocks is "OAuth-based login." The article treats this as a solved module and moves on. It is not solved. OAuth 2.0 was designed to delegate authorization from a human to an application. It was not designed for agent-to-agent delegation chains. Its gaps are well documented and, in my experience auditing contracts, brutally exploited: no fine-grained, composable permission credentials without bolt-on extensions like RAR, GNAP, or UCAN; systematic over-scoping, where tokens grant more than the task needs and revocation is coarse; no native way to express a constraint like "Agent A acts for User U, in scenario S, with a spend ceiling of X"; and no native support for delegation chains, where Agent A authorizes Agent B to authorize Agent C.
OAuth is the load-bearing assumption of PAP's identity layer, and it is the assumption most likely to fail under adversarial load. You don't build an agent identity system on a delegation model that was never designed for delegation chains. That's not a spec delay. That's an architectural fault, and no number of partner logos patches it.
When I audited Terra's oracle logic days before the collapse, I found the flaw and still didn't act on it, because confirmation bias told me the narrative was stronger than the code. The lesson I took from that $400,000 loss: find the load-bearing assumption, then assume it breaks. PAP's load-bearing assumption is OAuth. It breaks.
L2 has a resource-conflict problem the coverage doesn't touch. The .well-known URI space is scarce and mutually exclusive. If PAP claims /.well-known/poppy.json, A2A uses /.well-known/agent.json, and the MCP ecosystem uses its own paths, then any domain serving multiple agent protocols has to maintain multiple discovery files with overlapping semantics. This isn't a "standard hasn't shipped" problem. This is a discovery-layer coordination failure that persists even if PAP v0.1 ships on time. Two specs pointing at the same domain, both claiming to describe "the agent," is fragmentation baked into the design โ not fragmentation caused by a delay. The delay just gives the vendors cover to ship their own dialect first.
Now the memory taxonomy. Four types: session, semantic, procedural, episodic. Anyone who's read cognitive psychology recognizes this instantly. It's Tulving's semantic/episodic distinction plus procedural memory, relabeled for a product deck. When a vendor's architecture narrative maps one-to-one onto a 1972 academic taxonomy, you're looking at packaging, not engineering. The real problems โ storage tiering, cross-tenant isolation, memory-poisoning defenses, GDPR deletion rights across a persistent memory store โ none of them appear. Not in the article. Not in the vendor materials it cites.
And the orchestration claim: "multi-agent orchestration can run for hours to days." Read that carefully. "Hours to days" is a capability claim, not an engineering maturity proof. Long-running agents multiply three specific failure modes: hallucination accumulation, cost runaway, and permission drift. An agent running for seventy-two hours on a delegated OAuth token is seventy-two hours of unchecked scope. The article frames the runtime as a feature. It's a liability surface.
The coverage also ignores the one inference that actually matters for capital allocation. If agent orchestration runs for hours to days, and each agent needs persistent memory, then the compute profile shifts from bursty inference to sustained, stateful sessions. That's a different cost curve. It favors whoever can hold long-lived state cheaply โ which is not the same set of vendors that win on raw throughput. The article treats "hours to days" as a product feature. It's a datacenter bill, and nobody has priced it.
On valuation, the material gives me nothing. Zero funding figures, zero revenue, zero pricing for the commerce module. In a bear market, that's not a footnote โ it's the whole story. A protocol with no disclosed pricing has no business model, and a spec with no business model is a coordination experiment, not an investment. I'd rather see one number โ the fee the commerce layer takes per routed transaction โ than forty-one partner logos. The logos don't compound. The fee does.
There's also a source-quality problem I can't ignore. The coverage โ or the deconstruction around it โ calls Thomas Kurian "Google CEO." He's Google Cloud CEO. Sundar Pichai runs Alphabet. A title error at that altitude is a signature. It's what you get from second-hand sourcing or generated content. One error doesn't invalidate everything, but it tells you the technical fact-checking density is low. When the source can't get a CEO title right, discount every precise number in it by a notch โ including the six launch partners, the thirty-five design partners, the eighty-plus comments, and the four memory types. Specificity is not accuracy. It's often just the costume accuracy wears.
Everyone's watching the spec announcements. That's the retail read. Six partners on the sixth, forty-one on the ninth โ the headline number goes up, the engagement goes up, the narrative gets louder. Retail sees momentum. Retail sees a coalition forming.
Smart money sees the opposite. Forty-one partners in three days is not technical consensus. It's a press list. Technical consensus takes months. You argue about byte formats. You fight over trust anchors. You write the conformance suite. You fail the first three reference implementations. Then you ship v0.1. None of that shows up in a partner count. What shows up in a partner count is exclusivity recruitment. A logo wall is a fundraising artifact, not an engineering milestone.
Here's the tell nobody's pricing: there is no conformance suite. No test harness. No reference implementation published with a license and a repo address. Without conformance testing, "standard" means nothing. Every vendor implements the spec differently, and the standard fragments into N dialects the moment two vendors disagree. That's not a delayed standard. That's a standard that will never converge, because nothing forces convergence. A spec is a suggestion until a test suite makes it a constraint.
And that's exactly where the vendor extensions come in. The vacuum gets filled by whatever ships first, not by whatever is correct. Vendor extensions aren't filling the gap to be helpful. They're filling it to own the layer. Whoever defines the discovery path owns the registry. Whoever owns the registry owns the routing. Whoever owns the routing taxes every agent that touches it. This is the same playbook as the L2 wars. The real difference between OP Stack and ZK Stack was never the cryptography โ it was who could convince more projects to deploy chains first. Same dynamic here, one layer down. Whoever gets the most vendor extensions into production owns the agent stack, regardless of whose spec is technically superior.
I paid $400,000 to learn that lesson. Pain is just tuition; I paid in full so you don't. The lesson: when a coalition forms faster than a spec, you're not looking at infrastructure. You're looking at a land grab wearing infrastructure's clothes.
And the framing itself is selective. The "silence" narrative conveniently ignores what already shipped. MCP is the de facto tool-layer standard. A2A โ the protocol, not the registry โ already has adoption. Framing the whole stack as "three specs unshipped" erases the parts that work so the failure reads bigger. The story isn't "nothing shipped." The story is "the solved layers got solved, and the unsolved layers โ identity and trust โ are exactly the ones nobody can agree on." That's the real signal, and it's tradable.
I didn't come to this by reading press releases. I came to it by watching the same fragmentation eat the L2 market for three years. We don't trade the layer that's already solved. We trade the layer that's still contested.
Watch three things. None of them is the partner count.
One: does PAP v0.1 ship with an open license and a reference implementation repo? No repo, no adoption curve. A spec without a reference implementation is a press release with a schema.
Two: who signs the trust anchor for the A2A Trust-Authorization Overlay? If it's Google, the "open standard" is a single-vendor control plane with extra steps. If it's domain-holder self-signing, you get a trust layer with no root of trust. If it's a third-party CA, you get a new centralized dependency and a new attack surface. Whichever answer comes back, that's the layer where the next hundred billion in agent volume gets routed, taxed, and eventually exploited.
Three: is there a cross-protocol conformance suite? Not a spec. A test suite. If two vendors can both claim PAP compliance and still fail to interoperate, you don't have a standard. You have a brand.
The agents will ship. The products already have. What won't ship on schedule is the part that matters โ the identity and trust layer that decides whether an agent has the right to move your money. Until that converges, every "agent standard" is a vendor extension in a trench coat, and the stack fragments one integration at a time.
I don't trade the announcement. I trade the settlement. And nothing has settled yet.