As engineering teams deploy AI agents within Snowflake, the initial playbook is straightforward: create semantic views and point Cortex to the data. While this operational approach is incredibly effective for getting use cases off the ground quickly, it often prioritizes immediate execution over long-term architecture. As these agents move from prototypes to core enterprise tools, teams face a new challenge: how to manage and grow these systems when underlying business logic and data structures inevitably change.

This talk introduces a design approach that complements Snowflake Cortex Agents. Inspired by John McCarthy’s foundational "Advice Taker" concept (1959), we’ll explore how to move from tightly coupling an LLM to specific data paths, toward designing a more flexible, declarative environment. By treating the LLM as a reasoning frontend and the underlying tables of your data architecture as a deterministic backend, we can decouple reasoning from storage. Under this mental model, adapting an agent’s behavior honors McCarthy’s core rule: you don't rewrite the underlying program; you simply offer the system new advice by inserting new rows, rules, or context directly into your tables. Ultimately, we can apply this technique directly with Snowflake's semantic views, creating an ecosystem where data and knowledge are both equally open to this approach.

By separating state from reasoning, we can think differently about the boundary between our data and our models, ultimately expanding the footprint of what Cortex Agents can do. This framework offers a repeatable blueprint for an architecture of intent, giving teams a clear model to purposefully shape their agentic systems rather than just connecting components.