Serve many customers from one lakehouse without leaking between them: isolation models, Polaris grants, and tenant-scoped agent state.
Ship prompts, tools, models, and metric changes without breaking answers: agent manifests, shadow mode, canaries, and rollback.
Cache data agent results safely with canonical query keys, Iceberg snapshot freshness checks, and permission-aware cache keys.
Where small language models win inside a data agent and where they fail, plus a routing architecture that escalates only when needed.
