Ben Okonkwo: Jonathan, hey — so I had a genuinely strange Tuesday where I spent two hours reading a protocol spec for fun, and I'm not sure what that says about me.
Jonathan Ingles: It says you saw what I saw. MCP dropped something significant.
Ben Okonkwo: The July 28th update, yeah. And look, the thing that actually stopped me — David Soria Parra, who co-created this protocol and runs it at Anthropic, said the stateless transition is 'probably the biggest change we've ever made to the protocol.' That's the person who wrote it. That's not hype from a press release.
Jonathan Ingles: Frankly, that quote is the whole episode. Because what he's describing — ripping out the session layer entirely, the Mcp-Session-Id header, the initialize handshake, all of it — that's not a patch. That's a different protocol with the same name.
Ben Okonkwo: Hm — I'd push back slightly on 'different protocol,' but the architectural shift is real. And then AWS comes in same day with AgentCore Gateway support on Amazon Bedrock, which, I mean, same-day is not something you pull off without advance access to the spec.
Jonathan Ingles: No. You don't. The release candidate was public May 21st. AWS had two months.
Ben Okonkwo: Which is part of why the governance move matters — MCP is now officially under the Agentic AI Foundation, a Linux Foundation directed fund, with Anthropic and OpenAI both backing it. So the question isn't whether it's open. The question is whether 'open' means anything when the complexity landed somewhere else.
Jonathan Ingles: That's the thing we're pulling on today. Where did the complexity actually go.
Ben Okonkwo: Right — but complexity 'landing somewhere else' is still abstract. Let me make it concrete. Old MCP was like a phone call. You dial in, the line stays open, the other side remembers you because the connection is live. New MCP is like texting. Every message carries everything the recipient needs — who you are, what version you're running, what you can do — right there in a _meta field on each request. No open line.
Jonathan Ingles: That's the one. That's the whole architecture in one sentence.
Ben Okonkwo: And what that buys you practically is — now a server can run behind a standard round-robin load balancer. You don't need sticky sessions. You don't need a shared session store. Serverless, edge infrastructure, all of it suddenly works. That's real. That's not nothing.
Jonathan Ingles: Sure. But the RC was public May 21st. Two months before July 28th. So teams were already building against this spec.
Ben Okonkwo: Which means the 'landmark announcement' is actually a finalization of something already in the wild — and yeah, that matters for how we read the adoption numbers. The novelty is the official status, not the technical content.
Jonathan Ingles: The announcement was constructed. The surprise was manufactured.
Ben Okonkwo: Now — two extensions also graduated to official status in this release. MCP Apps, which is server-rendered UIs, and the Tasks extension, long-running asynchronous tasks. And that second one is interesting because, wait — actually that's the tell. If you've truly eliminated state at the session layer, why do you immediately need an official extension for long-running async tasks? State didn't disappear. It migrated.
Jonathan Ingles: Into the gateways. AWS AgentCore, Snowflake Cortex. That's where it went.
Ben Okonkwo: That's the question we need to sit with — because the spec got simpler, but someone is holding state for those long-running workflows, and right now that someone is the platform with the infrastructure budget to do it.
Jonathan Ingles: And that's the wrong take everyone's circulating right now — the narrative is that removing the session layer democratized agent building. I think it did the opposite. Here's the concrete version: imagine a solo developer, small logistics firm, reads 'simplified protocol, no session layer' and thinks the barrier dropped. Then they try to track a shipment across three async steps. The protocol gives them nothing. They need AWS AgentCore or Snowflake Cortex to hold that state. The spec got simpler. Their bill didn't.
Ben Okonkwo: Okay but — isn't running on serverless actually cheaper per-invocation than maintaining a persistent session store? The simple integration case genuinely did get easier.
Jonathan Ingles: I'll grant that. Fully. Simple tool calls, one-shot queries — yes, cheaper, easier. But production agent workflows always need state. And the features the spec removed — session management, state persistence, authentication orchestration — those are exactly the core value proposition of AgentCore and Cortex. That is not a coincidence.
Ben Okonkwo: Hm. So the spec didn't just simplify — it created the market for the gateways.
Jonathan Ingles: Designed the gap they fill, yeah.
Ben Okonkwo: Now — this is where the Runlayer and Rippling dispute comes in, because it's actually evidence of how high-stakes this already is. Runlayer accused Rippling of stealing its MCP product concept. A startup and an enterprise software company in a legal fight over gateway positioning — before the spec even had official status. That's not an ecosystem enthusiasm story. That's a land-grab.
Jonathan Ingles: Right — and that fight breaks out because everyone can see the gateway layer is where the money actually lives. Not the protocol. The protocol is open. The chokepoint isn't.
Ben Okonkwo: Which, I mean — wait, that's actually the assumption I want to flag. The one the whole 'democratization' claim rests on. It assumes the spec and the infrastructure layer have the same accessibility. They don't. And the part that I think makes this considerably worse is what happens when you ask who actually enforces anything at the Linux Foundation level — we'll get to that, but it complicates the 'open standard' framing significantly.
Jonathan Ingles: The protocol got simpler. The ecosystem got more centralized. That's the line.
Ben Okonkwo: And that centralization point — it rests on a claim that nobody has actually checked. The 400 million monthly SDK downloads figure. Anthropic and the Agentic AI Foundation put that number out. No npm forensics, no third-party telemetry, no independent repo traffic audit. And they're also citing 4x growth in 2026 alone, TypeScript and Python SDKs crossing a billion total downloads. These are the numbers that make 'industry standard' stick as a label.
Jonathan Ingles: Self-reported by the party with the most to gain from the number being large. That's not evidence. That's marketing with extra steps.
Ben Okonkwo: Right — but okay, what would actually count as verification? Because npm download data is public. Someone could cross-check the package integrity logs against what Anthropic is claiming. Nobody's done it, at least not publicly. That absence is its own data point.
Jonathan Ingles: The AWS same-day support — people read that as independent validation. It isn't.
Ben Okonkwo: No, I mean — Amazon Bedrock AgentCore had two months with the release candidate. And AWS has a direct commercial interest in being the default MCP gateway. That's not enthusiasm, that's positioning.
Jonathan Ingles: Which brings you to the governance question. The Linux Foundation structure, the Agentic AI Foundation — Anthropic and OpenAI both backers, direct competitors sitting at the same table. And David Soria Parra is still the lead maintainer. He's still Anthropic's hire. The protocol moved houses. The person who decides what goes in it didn't.
Ben Okonkwo: Hm. And the spec does have a formal 12-month deprecation policy now — that's a real structural constraint, not cosmetic. But 'real constraint' and 'enforceable constraint' are different things. Who sanctions a violation? The Linux Foundation doesn't have an enforcement mechanism that compels AWS or Anthropic to do anything.
Jonathan Ingles: Exactly. So what do you actually watch for?
Ben Okonkwo: Three things, concretely. One — does any independent team publish adoption data that either confirms or undercuts the 400 million figure. Two — does a smaller team, not AWS, not Snowflake, report that the stateless spec actually delivered the promised simplification end-to-end without a gateway. Three — does David Soria Parra's technical roadmap ever diverge from what Anthropic ships. If those three things stay unresolved, the 'open standard' framing is load-bearing on a foundation nobody's tested.
Jonathan Ingles: And none of those three questions point toward the protocol. They all point toward the platforms. Anthropic built MCP in November 2024 as an open wire — connect any agent, any system, no gatekeepers. The stateless spec genuinely achieved that. And then the complexity of running anything real migrated straight into AgentCore and Cortex. So the open wire exists. It just terminates inside someone's proprietary infrastructure.
Ben Okonkwo: The question I can't resolve is whether that was avoidable. Like — was there a spec design that kept complexity distributed? Or is this just what happens when you simplify a protocol that real workflows need state from? I genuinely don't know.
Jonathan Ingles: Neither do I. And I think that's the honest place to stop.
Ben Okonkwo: Good conversation. Uncomfortable in the right way.
Jonathan Ingles: Uncomfortable is the only kind worth having.
