I’ve been hearing the term “MCP” (Model Context Protocol) floating around a lot lately, especially after my company’s AI Week. I saw presentations and even a colleague demonstrating how simple it was to spin one up. I heard that MCP is easy to use, but I didn’t have a good grasp of what it was.

After watching a short, three-part video series, I understood what an MCP is. I now see why it’s needed and its role in the LLM ecosystem.

This led me to some discoveries. I learned about the “Agent” pattern and the “ReAct” (Reasoning and Acting) framework. This new knowledge cast a new light on a previous experience I had. I had been trying to use DeepSeek as a code generation model and noticed some platforms only offered “ask” or “edit” modes, with no “agent” capabilities. I had wondered if the model itself was the limitation. Now I realize that’s likely not the case. It’s more probable a strategic choice by the provider to steer users towards other models for agent-based tasks.

An analogy is: If the LLM is the CPU, then the Agent is the conductor. The LLM provides the raw processing power, but the Agent directs the symphony, coordinating tasks, and making decisions to achieve a complex goal.

The revelation around system prompts was a surprise. They can be very long and detailed for a sophisticated agent. When I think about the simple, ten-word system prompt I use for my calls on Cloudflare AI, the contrast is significant. It gave me a more realistic understanding of what “prompt engineering” truly entails. It’s not just about asking a clever question; it’s about architecting a detailed set of instructions and context to guide the model’s behavior.

I have a new understanding of LLMs, agents, and prompts.