-
What Goes in the Prompt, What Goes in the Tool
An agent keeps its knowledge in two places, and most broken agents have something on the wrong side. One question decides the split: does this change between calls? No: prompt. Yes: tool.
-
The Context Window Is a Desk, Not a Memory
The model holds nothing between your messages. Every turn, the entire conversation is packed onto one working surface and read from scratch. Once you see the desk, long-chat costs, sudden forgetting, and product memory all stop being mysterious.
-
It Never Knows That It Doesn't Know
There is no moment inside the model where it knows it does not know. The same machine that gives you a correct answer gives you a confident wrong one, and from the inside the two are identical. Hallucination is not a broken mode. It is the normal one, on thin ground.
-
It Only Ever Guesses the Next Word
The model does not decide what to say. It decides the next word, throws everything back in, and decides again. The paragraph you read as one thought was built one guess at a time. Once you see the loop, three things stop being mysterious.
-
The Model Never Reads Your Words
You type English. The model reads integers. It never sees a single letter of what you wrote. Understanding the thing in between, the token, is not trivia. It is the difference between guessing at cost, context, and prompt design, and knowing.
-
Don't Give Your Agent Root
MCP makes wiring an agent to your data trivial. That is exactly why the access decision is now the one that matters: what it can reach, what it can do, and why least privilege is the only rule that holds.
-
Agents Are Systems, Not Prompts
Everyone asks LangChain or CrewAI first. Wrong question. A production agent is architecture: a memory policy, a context budget, a model decision, and a prompt with guardrails. The framework is the easy part.
