<p>When AI agents fail on enterprise data, teams tend to blame the model. In practice, the root cause is usually upstream: a schema changed silently, a transformation behaved differently between development and production, or a data source was swapped without updating downstream assumptions. The data layer is the unreliable component, not the large language model (LLM).</p> <p>In this session, Tom Kaltofen examines why coding agents, such as Cursor and Claude Code, handle context reliably while enterprise data agents struggle. Coding agents operate on deterministic context: files have fixed paths, functions have type signatures, and tests provide ground truth. Enterprise data has none of that by default. From there, he introduces the handover layer pattern: data producers publish transformations with explicit input/output contracts, consumers declare what they need, and a runtime resolves how to compute it. The architecture follows the same principles as Kubernetes: declarative specs, independent controllers, and runtime reconciliation.</p> <p>A live demonstration shows an AI agent discovering transformations through a plugin registry, applying PII redaction, and switching between data sources without hard-coded paths. The demo runs on mloda, an open-source Python framework.</p> <p>The session covers what worked, what did not, and where the pattern breaks down. Attendees leave with a concrete architectural pattern for separating “what” to compute from “how” to compute it, and a clearer picture of why existing tooling, such as feature stores, semantic layers, and data catalogs, was not designed for AI agent consumers.</p>