Model Context Protocol Demystified

It is the age of AI – while AI is evolving many aspects of human interactions, the impact on software is particularly interesting. There is a huge opportunity for developers to infuse apps with solutions powered by generative AI and large/small language models – the end result should be smarter apps and better user experiences. Modern AI is also an opportunity to streamline and automate developer workflows for better productivity – AI can do much of the mundane code writing, freeing up developers to review code and guide architectural decisions.

Today’s generative AI Models are powerful systems capable of understanding human natural language to hold conversations and respond to prompts with text, code, audio, video and other types of creative content. Powered by unsupervised learning and deep neural networks, today’s general purpose AI models sport support for huge amounts of parameters/tokens and have massive processing power supported by seemingly unlimited wealth of knowledge.

However, despite all the smartness, AI Models struggle with a few foundational problems that are baked in by design:

• Knowledge used to train the models is often time stamped
• Grounding techniques like Retrieval Augmented Generation (RAG) are hard to scale
• AI Models are good at generative content, but cannot pull off concrete actions
• Models are pattern-matching systems, not computational workhorses
• AI lacks specific context for personal/enterprise tasks
• Without specilized tools, AI responses are not predictable & repeatable
• With randomness built-in, AI Models cannot be included in automation of tasks

As we graduate to more Agentic workflows where AI Agents perform concrete actions on behalf of users, context would be key – specialized tools that can provide grounding and help AI perform repeatable tasks.