This finally makes AI agents click

In this video, I break down AI agents into three building blocks: the model, tools, and retrieved context. Here’s the short recap.

TL;DR

• A model generates responses from the context it receives. It does not remember a conversation by itself: the application supplies the relevant history and instructions with each request.
• Your message is only part of the input. Custom instructions, project files such as AGENTS.md, and system instructions also shape the response. Everything has to fit within the context window.
• Tools let the model interact with the outside world. The model requests an action; the harness executes it and sends the result back so the model can continue. That loop is what makes an assistant agentic.
• MCP standardizes the connection between the harness and external tools. It does not make the model smarter, and a correctly formatted tool call can still use the wrong tool or arguments.
• RAG adds retrieved information—documents, a wiki, search results, or memory—to the model’s context. That information still consumes tokens, can become stale, and can help or hurt depending on its relevance.
• The harness matters as much as the model choice. Prompts, available tools, and the context you supply explain why the same model can behave differently in different products.

The takeaway: understanding these building blocks makes agents easier to reason about. Tools and retrieved context are pieces you compose for the task, rather than a single solution that fits everything.

Watch the full 17-minute video on YouTube.