42,753 Servers and a Dead Spec: The Agent Interop War Crypto Slept Through

I don care about the 42,753 number. Not the way everyone else does.

On 2026-10-13, the MCP Registry — Anthropic's tool-discovery layer for AI agents — crossed 42,753 registered servers. Every newsletter I opened that week led with it. Big green number. "Ecosystem thriving." "Developers voting with their code."

And sure, it's real. But here's the thing nobody put in the headline: the same week, the layer that actually decides whether two agents can trust each other — Google's A2A spec — was still sitting unratified. Not approved. Not rejected. Just stuck. Bleeding out quietly while Salesforce walked in through the side door.

I don trade on press releases. I trade on the gap between the number people celebrate and the number they're not counting. Right now that gap is the widest I've seen since I was manually tracing Parity transaction hashes at 3am in 2017, watching a "working" wallet bleed funds nobody had audited yet.

So let me be blunt. If you build anything in crypto — payments, agents, stablecoin rails — you need to understand what just happened in AI infrastructure this month. The war that went sideways over there is the exact war crypto keeps losing to itself.

Let me back up. Two protocols matter here, and they come from two different camps.

MCP — the Model Context Protocol — is Anthropic's. It standardizes how a language model connects to tools and data sources. Think of it as the plumbing that lets an agent say "go fetch this" without every developer writing a custom API wrapper from scratch. The MCP Registry is where those servers get listed and discovered. It works. It scaled. 42,753 servers is the proof.

A2A — Agent-to-Agent — is Google's. Different layer entirely. MCP handles tool discovery. A2A was supposed to handle the harder thing: agent-to-agent discovery and trust. How does Agent X find Agent Y, verify what Y claims it can do, and confirm Y is who it says it is? A2A was meant to answer that. Its spec laid out capabilities, permissions, identity.

And that spec — the one meant to give the whole industry a shared trust grammar — never got ratified.

Meanwhile, on 2026-10-12, Salesforce shipped Agentforce Third-Party Orchestration as part of Winter '27 GA. Its orchestration now extends to "any ecosystem that complies with A2A." Sounds open. Reads open. Is it?

That's the story. Not the 42,753. The fact that when the formal standard stalled, a single vendor's runtime quietly became the referee for what "compliant" even means.

This is crypto's oldest nightmare wearing new clothes. I've watched this movie. It's the same plot as every "open standard" that got eaten by whoever shipped first.

Let me draw the whole thing out, because the shape of it matters more than any single number.

[Tool discovery layer] MCP Registry — WORKS (42,753 servers, 2026-10-13) ↓ last-mile solved [Agent trust layer] A2A Registry — STALLED, governance vacuum ↓ pivot [Fallback path] AI Catalog + ARD — 2026-08 ↓ filled by [Vendor layer] Salesforce Agentforce — GA 2026-10-12

Two curves. One already crossed the tipping point. The other is stuck at the gate. And the reason isn't technical. It's governance.

When MCP hit scale, it proved the flywheel: clear, usable interface → builders adopt en masse → more servers → more reason to adopt. Textbook. But that same logic failed to replicate on A2A. The bottleneck was never feasibility. It was coordination. Getting a whole industry to agree on identity, permissions, and capability disclosure is a political problem dressed as an engineering one.

Compare that to how crypto governs its own standards. We've spent years learning that coordination is the hard part, not the code. DAO grant committees, foundation votes, the endless EIP debates — every one of those is the same political problem A2A just choked on. The difference is that crypto at least built mechanisms for it. Optimism's RetroPGF works because it funds outcomes after the fact instead of arguing about permission beforehand. A2A had no such mechanism. It had a spec and a hope.

Here's the weird part: there are no tokens in this fight. No airdrops, no governance coins, no staking. MCP's growth is pure utility-driven — developers adopt it because it saves them work, not because they're farming a subsidy. That's actually healthier than most crypto growth, and it's also why the governance gap is so stark. When there's no token to coordinate around, there's no built-in incentive for the crowd to force a standard through. You just get whoever ships first.

When the A2A spec hit the wall, the fallback was AI Catalog and ARD — the Agent Registry Definition. Read that pivot carefully. That's a standard downgrade. The ambition dropped from "industry-wide ratified spec" (de jure) to "a directory definition a vendor can actually implement" (de facto). From a shared law to a company's own house rules.

And that's where it gets interesting for anyone in crypto.

Because I've seen de jure collapse into de facto before. The 2017 break didn just teach me about multisig bugs — it taught me that when the formal process can't keep up, the market doesn't wait. It hands the pen to whoever's already shipping. Parity's contracts were "standard." Until they weren't, and nobody with authority had a fast answer.

Salesforce is that pen-holder now. Their orchestration "extends to any A2A-compliant ecosystem" — but they define the compliance. Their runtime is the judge. It's a walled garden with an open gate painted on it.

Look at who's positioned where. Salesforce sits downstream, closest to the money — enterprise buyers. It can define the standard by demand. MCP sits in the middle, already load-bearing, hard to rip out. A2A was supposed to sit in the trust layer and now sits in a drawer. When the middle of the stack is owned by one camp and the trust layer is owned by a vendor, the open spec isn't a competitor. It's a compatibility checkbox.

Now the part crypto keeps missing. The trust layer is the one that matters, and it's exactly the layer crypto knows how to build. A2A's original promise was capability and identity disclosure. Salesforce delivered the "capability" half as a self-published agent card — and explicitly no security or identity guarantee. Self-published, unsigned. That's a DID document with no signature verification. A claim, not a credential.

In crypto, we already fought this battle. We know an unsigned self-attestation is worthless the second someone wants to impersonate you. We built signing, verification, on-chain identity precisely because "trust me" doesn't scale. So the fact that the AI agent world is now running on unsigned self-published cards — while the trust layer sits vacant — is either a giant red flag or a giant opportunity. Usually both.

There's a second-order problem nobody's modeling yet. MCP comes from Anthropic's camp. A2A comes from Google's. If tool discovery lives in one registry and agent trust lives in another, you get a metadata interop problem — two registries that don't natively recognize each other's definitions. Builders will end up maintaining dual registrations, double the metadata, double the drift. That's friction that quietly rewards whoever offers a single front door. And a single front door is exactly what a walled garden is.

There's a hidden number inside the 42,753 too. Registry metrics lie. npm and PyPI taught us that — millions of packages, most abandoned. 42,753 could be registered servers. It is almost certainly not 42,753 active servers. I'd bet the active-to-registered ratio is ugly. Nobody's publishing it, which tells you something.

Here's what the crypto commentariat got wrong this month. They either ignored this story entirely, or filed it under "AI, not our problem."

Wrong on both counts.

First: the agent economy is going to run on crypto rails whether the standard-setters like it or not. Stablecoin payments are already the cheapest cross-border settlement layer on earth, and agents that transact will reach for them. If the agent interop layer locks into two or three incompatible walled gardens, crypto's payment rails get forced to plug into each garden separately — exactly the fragmentation stablecoins spent a decade escaping.

Second, and this is the contrarian bit: the A2A collapse isn't a failure to mourn. It's a vacancy to fill. The trust-and-identity layer is unoccupied. The formal spec died. The vendor replacement gives you connectivity without verification. That's a hole crypto is uniquely equipped to fill — verifiable identity, signed credentials, on-chain attestation.

Then there's the regulatory shadow. Walled gardens are the one thing regulators have historically hated — browsers, operating systems, app stores, all of it. If agent interop congeals into two or three incompatible gardens, the EU's Digital Markets Act playbook is sitting right there, gatekeeper rules and all. The irony is thick: the same Brussels that forced Apple to open its messaging is going to be staring down an agent layer that shipped closed by default and called itself open. Slow-moving risk, but the kind that rewrites the game once it lands.

I don think anyone in Brussels or San Francisco planned it this way. But the gap between "connectivity is abundant" and "trust is unverified" is the exact gap crypto identity infrastructure was built for. And it's open right now.

The trap is that builders get seduced by the connectivity. Salesforce makes integration easy today. Enterprise CIOs will close deals on Winter '27 before the standard question is ever resolved. Procurement is a vote. Once a thousand companies integrate through one vendor's runtime, migration cost locks the standard in place — and the open alternative isn't competing to win anymore, it's competing to be backward-compatible with a garden it never agreed to join.

That's the de facto trap. It doesn't announce itself. It just quietly becomes the default while everyone's counting servers.

So watch the ratio, not the number. Watch whether ARD is a real path or a polite exit. Watch whether the trust layer stays vacant long enough for a crypto-native identity standard to claim it.

The 2017 break didn make me slow. It made me watch the layer nobody's counting. That layer, right now, is agent trust. And it's wide open.

The cheetah's rule holds: the news that moves markets is never the headline. It's the line under it. Right now the line under it says the trust layer is unowned, and nobody's racing to claim it. That should bother you — or excite you, depending on which side of the trade you're standing on.

One question for the builders reading this: when your agent plugs into the "open" ecosystem this winter, do you actually know who's verifying the other side — or are you just trusting a card someone wrote about themselves?