Contrary to the summit's promotional framing, the number that matters out of MCP Dev Summit Toronto is not 500 million monthly SDK downloads. It is four. Four protocols. Four registries. Zero coordination standards binding them together. I spent the better part of last year dissecting the same structural failure in the AI-Crypto convergence stack โ specifically the Chainlink API surface, where I identified a flaw that let AI models manipulate price feeds through synthetic data injection. That was a single trust boundary. The agent economy now runs four, and no one owns the arbiter between them.
Here is what the summit actually disclosed, stripped of narrative fluff. MCP handles vertical tool access. A2A handles horizontal agent-to-agent communication โ and MCP explicitly does not cover it. AP2 handles payments. XAA handles identity. Each is mature on its own axis. None shares a policy or observability structure with the others. The Johns Hopkins workshop title said it outright: "Tool Integration versus Agent Communication." The "versus" is not stylistic. It is a confession.
To be precise about maturity, because marketing blurs it. MCP is at scale โ TELUS has run it in production for a year, Accenture claims 200,000 users. A2A is in production with 150-plus organizations under Linux Foundation stewardship. XAA is transitioning from proof-of-concept to production with 25-plus early adopters. The coordination layer is at none of these stages. It does not exist. MCP itself is built on JSON-RPC 2.0 over stdio and Streamable HTTP โ no computational novelty, only network effect. Its value is convergence, not breakthrough.
The download figure deserves its own dissection. 500 million monthly downloads is an SDK metric, not a developer count. CI/CD pipelines, container images, and dependency trees inflate it by orders of magnitude. A bug is just a feature that hasn't found its exploit yet โ and a download count is just a vanity metric that hasn't met a churn analysis. The summit quoted the number because it cannot quote production call volume, which would be far less flattering.
The technical gap has a name: a meta-protocol. MCP is vertical โ agent to tool. A2A is horizontal โ agent to agent. XAA is identity. AP2 is payment. They are orthogonal by design, and orthogonality without an orchestration standard means every integration point becomes a bespoke negotiation. There is no shared registry of capabilities, no cross-protocol capability discovery, no single place to answer what an agent is allowed to do, to whom, and for how much. Each answer lives in a different vendor's silo.
The tell is Google's Daryl Ducharme and his "One Router for Your Agent's MCP Servers." Read that as an engineer, not a marketer. When a protocol stack cannot manage itself, every implementer builds a single router. That is not a feature. That is a scar. The front-runner didn't solve the coordination problem โ it routed around it, on its own infrastructure, under its own control. Google now holds A2A, AP2, and the engineering workaround. That is the best-positioned seat at a table with no chair.
Now the incentive structure, which the summit's neutral tone carefully avoided. Anthropic open-sourced MCP and captures value only indirectly, through model and API calls. Okta's XAA is a subscription identity service โ it charges. AP2 touches payment rails. The layers that monetize sit above the layer that generated the ecosystem. This is value-distribution asymmetry dressed as interoperability. The missing coordination layer is not an accident of engineering. It is the rational outcome of four vendors each preferring their own layer become the de facto standard. Filling the gap is nobody's interest. That is why it stays empty.
The registries are the quietest story. 15,930 public servers spread across four directories, none federated. A registry is a discovery layer, and a discovery layer is a trust layer, and a trust layer is a toll point. Whoever federates those four registries โ or replaces them with one โ owns agent discovery the way ICANN owns DNS. No vendor has proposed it. Every vendor benefits from the fragmentation.
The commercial signal everyone cites โ Cisco's RSA data showing 85% of enterprises in trial and 5% in production โ is being read as a coordination-layer deficit. I am less certain. It could equally mean enterprises have not found agent use cases worth the integration cost. The summit attributes 80 percentage points of unrealized conversion to one missing standard. That is convenient. It is also over-attributed. Correlation with a gap is not proof the gap is the cause.
Security is where this stops being academic. Every added protocol layer is an added trust boundary, and there is no meta-layer enforcing policy across them. Prompt injection propagates through MCP tool calls with no unified filter. XAA grants agents independent identities โ solving who an agent acts as, while creating lateral-movement risk when an agent identity is stolen and can reach every application the user touches. AP2 lets agents initiate payments. Inject that, and the loss is capital, not just data. CrowdStrike dedicated a session to insider threat in the agent stack, which is a security vendor telling you where the next attack surface lives.
The coordination gap is simultaneously a technical hole, a commercial tollbooth, and a security failure โ and the three cannot be solved independently. Unified policy enforcement is the same artifact as unified observability, which is the same artifact as unified audit logging. My work on Trustless AI Oracles โ the zero-knowledge framework cited in the EU's AI Act guidelines โ reached the same conclusion from a different direction: verification without a shared enforcement point is theater. The EU's AI Act does not wait for protocol communities to reach consensus. It imposes obligations on deployers regardless of whether a coordination standard exists. That is the pressure that manufactures one โ not goodwill, not open-source idealism, but a compliance deadline with penalties attached.
Which brings me to what the bulls got right, because they got something right. The standards genuinely reduce integration cost. MCP servers are replacing hand-rolled API adapters; A2A is replacing point-to-point integration; XAA is replacing traditional SSO for a new class of non-human principal. The network effects are real โ 15,930 public servers across four registries is a genuine ecosystem, not a slide. And Linux Foundation stewardship of A2A signals the industry has internalized the lesson that single-vendor protocol control destroys trust. If a coordination standard does emerge, the neutral-governance precedent already exists. I will concede that.
What I will not concede is the summit's implied resolution: that the protocol community self-organizes its way to a coordination layer. Based on my audit experience, the layer that binds MCP, A2A, XAA, and AP2 will not be a protocol. It will be a product. Runtime platforms โ AWS, Azure, GCP agent hosting, or orchestration frameworks like LangGraph and Temporal โ already own the execution environment where cross-protocol policy would live. They will ship coordination as a managed feature, bundle it with compute, and quietly re-monopolize the openness MCP was built to preserve.
Watch three things. Whether a formal coordination-layer proposal lands at the Linux Foundation or IETF in the second half of 2026 โ and who authors it. Whether Okta's Agent SSO adoption after its August 2026 launch comes with a security incident that retroactively prices the identity risk. And whether cloud vendors' agent platforms become the de facto orchestration layer, making the standards question moot. The 85% to 5% convergence will not be measured by keynotes. It will be measured by whose tollbooth every agent transaction passes through.
The narrative survived the summit. The mechanism didn't. The agent economy is about to discover it has no integrity layer at the coordination boundary โ and the vendors are counting on that discovery arriving too late.