Who Connects the Machines?

Two weeks ago, we wrote about the diffusion curve of AI. The idea was simple: adoption is still concentrated among clusters of power users, and the more interesting question is not whether people support AI, but how quickly it becomes embedded across different kinds of work.

Last week, we argued that better models are only part of the story. The real accelerant is the harness around the model. Memory, tools, scheduling, and persistence are what transform a chatbot into something that can actually perform useful work.

Because once you have a capable model and a capable harness, another question emerges: how does it connect to everything else?

The answer sounds technical, but it may end up being one of the most important developments in AI over the next few years. We are entering what looks increasingly like a protocol layer.

Most technological revolutions eventually run into the same constraint. The technology itself works, but the ecosystem around it is fragmented. Every platform speaks a different language. Every tool requires a different integration. Every company builds its own connectors. Progress happens, but it happens inefficiently, and the cost of connecting systems becomes a limiting factor on how quickly the technology spreads.

The internet faced this problem early on. Networks existed, but they did not reliably interoperate. The breakthrough was not just better machines, but shared standards that defined how machines should communicate. TCP/IP standardized how data moves between systems. HTTP standardized how information is structured and requested on the web. Once those standards emerged, the internet stopped being a set of isolated systems and became a coherent platform that others could build on.

AI is beginning to run into a similar constraint.

Today, if you want an agent to access documents, search across tools, update internal systems, query databases, or send messages, each connection has to be built explicitly. Historically, every model provider, application, and agent framework solved this in its own way. The result is duplication of effort and a growing maintenance burden as the number of tools increases.

What makes the current moment interesting is that the industry is beginning to converge on shared approaches to this problem.

The most important of these is Model Context Protocol, or MCP, introduced by Anthropic in 2024. At its core, MCP is a standardized way for AI systems to connect to external tools and data sources. Instead of building custom integrations for every model and every application, developers can build a single interface that any compatible system can use.

On the surface, this looks like a developer convenience. In practice, it is a structural change in how capability is distributed across the ecosystem. A protocol turns integration from a series of bespoke engineering projects into shared infrastructure. Once a tool exposes an MCP interface, any compliant model or agent can potentially interact with it without additional work.

That idea has started to spread quickly. OpenAI added support for MCP. Microsoft integrated it into parts of its ecosystem. Open-source frameworks began building around it, and a growing set of community-maintained servers now connect models to everything from databases and file systems to internal business tools and external services.

The important dynamic is not any single implementation, but the accumulation of integrations. Each new connection makes the standard more useful. Each new tool increases the value of adopting the same interface. Over time, this creates a compounding effect where the cost of not participating in the standard increases relative to joining it.

At the same time, other approaches are emerging that target adjacent problems in the stack. Some focus on how agents communicate with each other. Others focus on how workflows are orchestrated across multiple systems. But MCP has so far become the most widely referenced attempt at solving the core problem of tool access and interoperability.

This is where the story connects back to the previous two weeks.

If diffusion depends on practical usefulness, and usefulness depends on harnesses, then harnesses ultimately depend on connectivity. An agent that cannot reliably interact with the systems where work happens is constrained, regardless of how capable the underlying model is or how sophisticated the orchestration layer becomes. The real value of agents emerges only when they can move across the same surfaces where human work already exists.

That is what makes protocols structurally important. They sit one layer below applications and one layer below harnesses. They are not trying to improve a specific tool or workflow. They are trying to define how all tools and workflows connect.

Historically, these layers tend to matter more than they first appear. Some of the most durable companies in technology were not those that built the most visible applications, but those that controlled the interfaces beneath them. Operating systems defined how software ran on personal computers. Payment networks defined how value moved between institutions. Cloud platforms defined how infrastructure was consumed. Internet protocols defined how systems communicated at scale.

In each case, the protocol layer became more valuable as the ecosystem around it expanded.

Whether MCP becomes that layer in AI is not guaranteed. Today, a lot of agent work still runs through CLIs, SDKs, and direct API calls. Standards rarely win on technical merit alone. They win when enough people converge on them for practical reasons.

Most users will never notice any of this directly. Nobody thinks about HTTP when they open https://www.amazon.com/ to buy their kid’s school supplies or TCP/IP when they send an email or a WhatsApp message to family across the world. The most successful protocols tend to disappear precisely because they work. They become invisible infrastructure that everything else depends on.

While attention remains focused on model performance and visible product releases, a quieter shift is taking place underneath. The next phase of AI adoption may depend less on incremental improvements in intelligence and more on how easily that intelligence can be connected to the systems where real work happens.

If that is true, then the most important competition in AI may not be happening at the model layer. It may be happening one level below it, where the standards that determine how everything connects are still being written.