MCPs: Agentic Interfaces Are Getting Defined

"Define the protobuf first" - usually the first step of any design at Google. Protobufs (protocol buffers) at Google are the lingua franca of system communication. This seemingly a trivial rule actually has a lot of wisdom baked into it. What are the summary of key changes? Look at the protobuf. How will this affect downstream users? Look at the protobuf.... it's all about the interfaces at the end of the day. But what happens if we have a poorly designed interface? 

Natural Language: A Double-Edged Sword

What is the lingua franca for humans? Depends on the population you have! If you take a group of English speakers would they all understand each other? I struggle sometimes to understand things I've said before, there's dialects, region nuances. Flexibility is what you get with a language, but it comes with the risk of misunderstanding. With a few jumps we can see how this all ties back to Shannon's A Mathematical Theory of Communication. The more flexible an way of communication, the less "information dense" it is. Normal conversation is less precise than jargon, jargon is less precise than a communication standard, ....., natural language is less precise than a standardized interface....

LLMs have grown in their popularity because they ARE flexible, but we're still figuring out how to deliver high quality outputs reliably on a repeatible basis.

MCPs - Let's Add Some Structure to AI

Model Context Protocol (MCP) was introduced late last year by Anthropic and opens the door to a world of possibilities. It short is a standard for how to connect an LLM to data or service. Seems easy! ... and it kind of is! There's plenty of sites out there offering MCP services for common APIs like Google Maps, Github, etc.

The key breakthrough to note is we can now reliably connect strcutred systems to flexible LLMs.

If you're a product manager looking to create new personalization experience, you just need an MCP server for your existing data stores. If you have tons of microservices that you want LLMs to use, build an MCP! MCPs are the abstraction between a tool and an LLM.

Agency with MCP

There's quite a few definitions of what it means to be "Agentic" but it seems fundamental to have at least these three components: 

• Tool Use - The ability to operate within an environment means being able to ingest data and make requests for other services
• Memory - Having context of prior conversations, interactions, or events to enable continuity of an experience
• Decision Making - The abiliy to evaluation options and take autonomous actions aligned with a goal or a prompt.

MCPs provide us a path forward to well defined Tool Use. As Open AI, Anthropic, and others continue to adopt them, MCPs could become the standard for enabling agent interaction with external services.

Products with MCP 

From a PM's perspective, here's what MCPs mean: 

• MCPs provide a guideline for data & tool re-use in the generative AI era
• MCPs help achieve more reproducible behavior when providing context to an LLM
• MCPs are the on-ramp to agentic systems that can reason, remember, and act