Middleware Tries to Make Companies That Don't Want to be Dumb Pipes into Dumb Pipes

The lessons that people who love MCP need to learn from Web 2.0

So, OpenAI is backing Anthropic's MCP standard. The idea? A common language so AI agents can actually do stuff by talking to the apps we use every day and plug them together – hail an Uber, order groceries via Instacart, update Salesforce. It's classic middleware: a layer to smooth over the differences between systems. Developers always crave this kind of abstraction; it promises simplicity and power. Connect once, talk to many. But history suggests this dream often hits two very hard walls.

The first wall is technical: the 'lowest common denominator' problem. You're trying to create a universal translator for wildly different, complex beasts. Instacart's logic for substitutions and delivery windows has nothing in common with Salesforce's intricate object relationships or Uber's dynamic pricing. Any standard trying to cover them all inevitably sands off the specific features and nuances that make each service valuable. It becomes a blunt instrument. Abstractions often fail to capture the richness and performance of the native thing. Can MCP avoid simplifying its connected services into near uselessness?

Maybe yes! Large Lange Models really do allow us to fuzzily plug things together that the more-rigid classic API plumbing couldn’t. We get more things fuzzily plugged together but we can plug them together less reliably.

The second wall is strategic, and maybe insurmountable: incentives. Why would Instacart want to become a faceless backend for OpenAI's (or Anthropic's, or Google's) grand AI ambitions? Instacart's money isn't just in groceries; it's in ads within their app, controlling the user's discovery and shopping experience. Uber wants to nudge you towards Uber Black or sell you a subscription. Salesforce wants you embedded deep in their ecosystem, using their AI tools. Handing over the user interaction layer to a third-party AI means losing control over monetization, customer relationship, and strategic direction. They risk becoming commoditized pipes for someone else's empire. Nobody with have a business brain signs up for that.

We saw this play out in Web 2.0. Remember when Twitter, LinkedIn, Facebook, and others had vibrant API ecosystems? Developers built cool things on top. Then, as those platforms consolidated power, they choked off or severely limited that API access. Why? Because they realized the value was in owning the user experience and the data stream directly, not enabling others to build potentially competing experiences.

The tech ecosystem consolidated must birth a Cambrian explosion on top. The Cambrian explosion on top must consolidate.

But there’s one big example where this didn’t happen: Google Search and The Web (for a long time at least). What’s different?

Middleware (like MCP) or standardized formats (like HTML) thrive only when the components being connected remain relatively decentralized, and the centralization occurs via an external discovery mechanism – think Google indexing the open web. When the things-that-the-middleware-plugs-together becomes the platform, they become the centralizing force, and they eventually lock it down. Facebook (Meta) is the biggest example of this. A very clever friend once described Facebook as “The internet on the internet” for this reason.

So, while the idea of AI agents seamlessly orchestrating tasks across services via MCP is compelling, the structural challenges are immense. Can a standard overcome the inherent complexity without becoming uselessly generic? And more critically, can it align the incentives of platforms whose business models depend on owning the user, not being an invisible API call? History suggests probably not.

What kind of radically different incentive structure might actually convince platforms like Instacart or Uber to fully embrace being a 'dumb pipe' for external AI agents? Or maybe an entirely new class of entities will come into existence instead? Maybe a wave of hundreds of thousands of small companies with less than 5 people can be plugged into a broader ecosystem via MCPs.