Coding agents succeed because they have structured, static context, unlike organizational agents which struggle with dynamic data, governance, and evaluation.
Nine core data management concepts, from feature stores to retrieval layers, are explained using a series of relatable family and travel analogies.
An abstraction layer using a plugin system is proposed to orchestrate various data tools, creating composable and shareable data pipelines for AI agents.
A command-line demo shows an AI agent dynamically using and comparing different plugins for data redaction and tracing data lineage.
Using stable plugin interfaces brings software engineering best practices like testing and validation to data, making agent context reproducible and scalable.
Jobs that call for the skills explored in this talk.