Treat your context like a dependency. Your AI agents need to import context just like you import a library. If they are relying on stale metadata or "thin" documentation, they will fail silently. If your AI agent cannot "import" the current state of your business logic, it is already broken. When you treat context as static text (system prompts, wikis, PDFs), you introduce a silent failure mode. The data team might update the definition of a domain entity, the agent would still be using the hardcoded definition from last month's prompt. The fix is to treat context like code. 𝗧𝗵𝗶𝗻𝗸 𝗼𝗳 𝘁𝗵𝗲 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 𝗮𝘀 𝗮 𝗱𝗲𝘀𝗶𝗴𝗻 𝗽𝗮𝘁𝘁𝗲𝗿𝗻. For example, you can implement this by configuring @getdbt to act as an MCP server, which effectively turns your data models and metadata into a service consumable by AI. This setup opens up the semantic layer and discovery API to external tools, and enables features like governed text-to-SQL and direct query execution. You get a standardized pipeline where clients (like Cursor, Claude, or a custom agent) can "import" your dbt project's logic, and ensure all context is both auditable and enforceable . 𝗩𝗲𝗿𝘀𝗶𝗼𝗻 𝘆𝗼𝘂𝗿 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 Your semantic definitions should live in a repository, not a document. 𝗦𝗲𝗿𝘃𝗲 𝗶𝘁 𝘃𝗶𝗮 𝗔𝗣𝗜 Use standards like the Model Context Protocol (MCP) to let agents fetch the latest context dynamically. 𝗙𝗮𝗶𝗹 𝗼𝗻 "𝗜𝗺𝗽𝗼𝗿𝘁 𝗘𝗿𝗿𝗼𝗿" If the agent can't verify the freshness or ownership of the data it’s accessing, it shouldn't answer. I've collaborated with @getdbt to highlight a really fascinating report they just released on this architectural shift. It breaks down the pattern in detail: fandf.co/4adyPww #DataEngineering #AI #dbt #MCP #DevOps