Unlocking a new era of Agentic Expertise

Todays models are already remarkably "intelligent" meaning their training data contains most of the publicly available and some copyrighted knowledge. Making them capable of solving complex problems.

But most day to day tasks require understanding domain specific knowledge such as company specific data structures, APIs, weird legacy workarounds or internal processes or sector specialities.

“Context is worth 80 IQ points.” ~ Alan Kay

Knowing that we are working with an extremely capable apprentice we know that we just need to let them know where to find the necessary information. This already works well with web_search tools or RAG pipelines. But only to a certain extent. Plain information combined with an smart brain receives a reality check in the real world, where sadly things aren't always logical and make sense. So we need to aid plain information with context in which to place it.

But first we have to ask ourselves what is actually needed to complete tasks?

1. General knowledge / intelligence
2. Domain specific knowledge
3. Knowledge of when and which references to use to retrieve missing information / context
4. Equipment to carry out the task

Since we have already established that LLMs tick the first box, we can focus on handing out equipment for retrieving unfamiliar information and executing tasks. This is where the Model Context Protocol (MCP) comes in. MCP is like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect your devices to various peripherals and accessories, MCP provides a standardized way to connect AI models to different data sources and tools.

Enabling the future master of a craft by providing access to the tools of the trade.

With MCP we have a standardized way of supplying the tools to LLMs, meaning we can give them access to softwares to interact with real world data in documents, spreadsheets, databases, etc. they can retrieve, update, create and delete information. This marks a change from text in text out chat bots to actual agents capable of carrying out tasks. And a step from AI into the "real world".

The meaning of any event is not determined by the event itself, but by the context in which it is placed.

If you want to go deeper on the MCP protocol, I have written a compendium about it.