Hook
The keynote lineup for AGNTCon+MCPCon, scheduled for October 2026, has been circulating for weeks as though it were an industry scoreboard. Three MCP keynotes. Zero A2A keynotes. Three Google slots. One AWS slot. Two Solo.io slots. Traders I respect are already pricing "winners" and "losers" off that distribution, rotating exposure on the strength of a stage schedule. That is a category error, and it is the kind that quietly empties a position while you are still congratulating yourself on the read.
Charts lie. Intuition speaks. An agenda is neither a chart nor a forecast — it is a receipt. What the parsed agenda analysis actually hands us is a budget document wearing the costume of a technology map. The single most informative line in the entire program is the one the analysis never foregrounds: Solo.io, which appears on the main stage with both its CEO and its CPO, is also listed as an event sponsor. Once you see that, the "strategic map" framing collapses into something far more mundane — and far more useful.
Context
To see why, you need to know what these protocols are, and — more importantly — what they are not.
MCP, the Model Context Protocol, originated at Anthropic. It is a client-server specification that standardizes how a language model connects to external tools and data sources. It is an interface contract for tool access, nothing more dramatic. A2A, the Agent2Agent protocol, was initiated by Google and has since been donated to the Linux Foundation. It governs interoperability between agents — how one agent delegates a task, how agents communicate, negotiate, and hand off state. agentgateway, from Solo.io, is a protocol proxy: traffic management for agent-to-tool and agent-to-agent flows. That is infrastructure engineering, not architecture.
None of these is a control plane. The genuine control planes — with strict cloud-native definitions — are the vendor orchestration platforms: Vertex AI Agent Engine, Bedrock AgentCore, Azure AI Foundry. In cloud-native vocabulary, the control plane makes global decisions about the data plane; it owns reconciliation and desired state. MCP does not make decisions. It describes capabilities. Conflating the two inflates the technical stakes and smuggles in a narrative the evidence cannot carry. That conceptual slippage runs through the entire analysis, and it is the key to reading it: the writer is a cloud-native person — service mesh, API gateway, CNCF — and the worldview leaks. When your only tool is a gateway, every protocol starts to look like a control plane.
The protocol layer matters because it is where lock-in is either prevented or installed. Whoever controls the default interface controls the migration cost, and migration cost is a moat measured in engineer-hours.
The event itself is a two-track conference positioned at the intersection of agent frameworks and the AI-crypto convergence I have been trading since early 2026. That convergence is not abstract for me. When I run sentiment-augmented strategies across autonomous agent protocols, I am depending on exactly these interfaces — the tool-access layer, the delegation layer, the gateway in between. So the agenda is not trivia. It is the substrate under my positions, and I read it the way I read a contract: for what it promises and for what it hides.
Core
Let me build this like code: premise, logic, execution, result.
Premise: keynote allocation reflects marketing priority and sponsorship budget. Logic: conferences allocate stage time in proportion to sponsor contribution and organizer preference, not to deployment metrics. Execution: the evidence supports this cleanly. Result: reading competitive outcomes from seat counts is invalid.
Start with MCP, because it holds the strongest genuine signal — one the analysis underplays. MCP occupies three keynotes, and one of them is delivered by Thomas Sarlandie of OpenAI under the title "MCP in the Wild." That is the single most information-dense fact in the program. It means MCP has migrated from a private Anthropic asset into a multi-vendor de facto standard. OpenAI — a company that publicly backed MCP in 2025 — is now evangelizing it from the main stage. Cross-vendor adoption is the real scoreboard, and here it reads cleanly: MCP is winning the tool-and-context layer, and it is winning by becoming boring, which is how standards actually win.
A2A has zero keynotes and only a workshop. The lazy reading is that A2A lost. That reading is wrong. A2A and MCP are complementary more than competitive — MCP handles tool and context access, A2A handles agent-to-agent communication. And A2A's Linux Foundation governance is precisely the precondition for enterprise adoption in regulated sectors. A downgrade from keynote to workshop reflects stage economics, not architectural defeat. Code doesn't lie; a demoted session slot does not overwrite a governance advantage.
Google's three slots describe an unmistakable full-stack ambition: Rao Surapaneni on cloud platform governance and observability, Sarah Drasner on the Agentic Web, and Chi Wang — founder of AutoGen — on his personal open-source agent, Sutando. That spans cloud-scale architecture, network paradigm, and personal endpoint. It is a lifecycle play, from datacenter to pocket, and it is the most coherent strategy on the board.
AWS takes a single, deliberately narrow slot: Shadaj Laddad on deterministic simulation and universal properties. That is differentiation by subtraction — positioning as the verifiable substrate for high-risk agent operations, sidestepping the platform-governance red ocean. It tells you where AWS believes the defensible moat sits: correctness, not coordination. I have argued for years that the moat in infrastructure is never the feature list; it is the failure mode you can prove you handle.
Solo.io plays the governance card. Its keynote lands immediately after it donated agentgateway to the Linux Foundation. Donate the open core, buy neutrality and a seat at the standard-setting table. Classic Open Core plus standard capture. But the sponsorship disclosure undercuts the "seats equal technical leadership" reading entirely. The agenda is a funnel, and Solo.io is buying the top of it.
Now the commercialization layer, where the money actually lives. Value capture in agent infrastructure will stratify. Cloud vendors monetize through consumption — the compute and API usage generated when agents run on their platforms, the thickest margin in the stack. Gateway and protocol tooling vendors monetize through Open Core enterprise editions. The protocol standards themselves monetize almost nothing directly. MCP is open; Anthropic cannot charge for it. What it buys is ecosystem gravity — developers, servers, integrators clustering around the standard. A2A's neutrality buys enterprise trust. The conference is not a strategy map; it is the visible top of a go-to-market funnel, and the keynotes are lead generation.
This is the public-goods trap of protocol layers. The more open the standard, the harder direct monetization becomes, and the more value migrates toward cloud consumption and enterprise value-added services. Expect a land-grab of vertical MCP server startups — connectors, integrations, vertical tooling — followed by platform vendors absorbing them natively. Owning a standard is not owning revenue. The analysis treats "standard as moat" as an axiom; the axiom is unpriced.
On the technical route, one correction matters: MCP and A2A are engineering-level and compositional-level innovation, not architectural breakthroughs. They reframe existing RPC, API, and messaging patterns for an agent context. That is valuable, and it is not a new compute paradigm. The framing inflates them; the engineering does not need the inflation.
Industry impact: if MCP consolidates as a cross-vendor standard, ecosystem fragmentation risk drops, and developers integrate once and reuse across platforms. Traditional integration and API management vendors — the MuleSofts, Kongs, and Apigees — face reconstruction pressure as the agent-gateway category emerges. Historical rhythm suggests standard races resolve within twelve to twenty-four months. Kubernetes and OpenAPI set that tempo, and neither was settled by a keynote count.
Contrarian
Here is what the agenda analysis refuses to confront.
Microsoft and Meta are absent. Microsoft owns AutoGen, Semantic Kernel, and Copilot — it is among the largest players in the agent space. If this agenda were a true strategic map, that absence would be disqualifying. It is not a strategic map. It is an agenda shaped by sponsorship and organizer preference, and the missing names prove it. The more telling signal is buried: Chi Wang, AutoGen's founder, now appears under Google DeepMind. If accurate, that hints at an exodus of AutoGen ecosystem leadership away from Microsoft — a talent-flow signal the analysis never names, even though it is sitting in plain sight on the program.
Then there is the concept laundering. Wrapping protocol standardization as a "control plane war" inflates complexity to raise narrative tension. It also buries the actual technical risk. The real unsolved problems in agent protocols are not who owns the standard. They are security boundaries, permission models, tool poisoning, and context injection. A protocol that standardizes tool access also standardizes the attack surface. I spent 2022 funding independent security reviews and finding reentrancy bugs in mid-cap L2s. The lesson transfers directly: the interface that connects systems is the interface that leaks. Nobody on that stage is keynoting about minimum-privilege tool invocation or audit trails — and that silence is more consequential than any seat count. The code is the risk.
Watch, too, for the tension between Sutando's local-first, privacy-prioritized agent and the cloud vendors' orchestration. That is a genuine architectural fork — a decentralization play against the gravitational pull of the cloud — and the analysis waves at it without opening it. That fork, not the keynote ledger, is the story worth following.
Takeaway
Stop pricing protocols off keynote distribution. Price them off deployment metrics: MCP server counts, enterprise A2A integrations, agent platform pricing models — per token, per agent, per seat — and gross-margin structure. Those numbers, not stage time, will settle the standard war.
The question is not who got the keynote. The question is who owns the permission model when an autonomous agent signs a transaction. Charts lie. Intuition speaks. The seat count might be the loudest thing in the room — and the least true.