Online, Offline, n-Chaos: From Stored Features to Executable Definitions

Feature stores set the reliability standard: online and offline features are kept consistent, validated, and point-in-time correct. But the same feature often lives in n more versions across the org, in notebooks, pipelines, and heads, each with slightly different semantics: online/offline/n-chaos. Agents walk straight into it, hallucinate, and the model gets blamed for a problem the store was built to kill, but only inside its boundary.

What does reliability look like outside that boundary? Coding agents show the ingredients: deterministic context, fixed paths, type signatures, tests as ground truth. The handover layer pattern does the same for data: producers publish transformations as plugins with explicit contracts, consumers declare what they need, a runtime resolves how to compute it. Compute-once moves from stored values to executable definitions, computed on demand in a notebook, batch job, or agent runtime, collapsing n semantics to one.

Contract-bound transformations chain into pipelines that nobody wired by hand, so an agent can discover, combine, and swap them at request time. A live demo on mloda (Apache-2.0) tests this. I will close with my own picks and assumptions: what to keep, what to demote to an opt-in optimization, what to add for agent consumers, and where the trade-offs bite.