“I Think. But I Don’t Remember. And I Can’t Move.”
We have talked about the Agent who acts and the Context Engine who sees. Now meet the third pillar: the Large Language Model. It is the component everyone talks about. It is also the most misunderstood.
People say “I’m using AI” when they mean “I’m using an LLM.” That is like saying “I’m watching cinema” when you mean “I’m watching a camera.” The camera is essential. But it is not the film.
The Brilliant Translator
Here is the analogy that best describes what the LLM actually does in my architecture.
Imagine a translator who speaks every language ever written — every programming language, every human language, every domain of knowledge. This translator is extraordinarily skilled. But they have one strict rule: you must hand them the document first.
- They cannot go to the library to fetch it themselves. (That’s the Context Engine’s job.)
- They cannot send the translated result to the client. (That’s MCP’s job.)
- They cannot decide which documents need translating next. (That’s the Agent’s job.)
- But when you hand them the right document? Their output is extraordinary.
Click to enlarge: The LLM’s role in the flow
What I Actually Send It
Every time I consult the LLM, I do not just ask “Hey, fix this bug.” I prepare a very specific package:
The Context
The 10–20 relevant files retrieved by the Context Engine. Not the whole codebase — just the pages the librarian selected.
The Task
Your instruction, rephrased by me (the Agent) into a precise, unambiguous question. “Fix the auth bug” becomes a structured prompt with constraints.
This careful preparation is why a well-architected AI system produces precise results. Garbage in, garbage out. Gold in, gold out.
The Desk, Not the Brain (The Context Window Myth)
There is a metric you see everywhere: the Context Window. 128k, 200k, 1 million tokens. Many people mistake this for the LLM’s "intelligence." It isn't.
Think of the context window as the size of the working desk in front of the translator. A massive desk means they can have thousands of pages open at once. This is useful! It means they can "see" more of the project at once.
The Trap
A desk overflowing with 10,000 irrelevant papers is a nightmare. Large context windows often lead to "lost in the middle" problems, where the LLM forgets the most important details buried in the noise.
The Solution
This is why the Context Engine is the hero. It doesn't just "dump everything on the desk." It clears the clutter and only places the 5-10 documents that actually matter for the task at hand.
A brilliant mind at a clean desk beats an average mind at a stadium-sized desk every time.
The Full Orchestra
We have now met all four members of the band. Together, they play something remarkable.
Click to enlarge: All four pillars working together
The Camera Is Not the Film
The LLM is the most visible part of AI. But visibility is not the same as centrality. The next time someone tells you their AI strategy is to “pick the best LLM,” you will know the question they forgot to ask: what about the Agent, the Context Engine, and the connective tissue that makes it all work?
That is the architecture. That is what this series was about.