Okta Shares How It Governs Context for AI Agents at Scale

At dbt Summit 2026, Okta analytics engineering leader Pooja Crahen described how her team tackled the trust problem behind AI agents. Her roughly thirteen-person team, embedded across delivery, go-to-market, product, and finance, works within a technology, data, and intelligence organization running Tableau for reporting and Snowflake as its central warehouse.

Okta's early AI experiments used semantic views built inside Snowflake's UI and connected to Snowflake Cortex agents. They worked quickly, but sat apart from the governed dbt models defining business logic. Crahen's team identified three cracks: context fragmentation between Snowflake and dbt, no auditability beyond UI click history, and agent instructions flattened into one undifferentiated blob.

The fix was a three-layer agent context framework built in dbt and shipped through pull requests. Model and column metadata gained an "agent" sub-block in YAML; markdown context docs capture company-wide facts and routing rules; and semantic views moved into dbt's semantic layer spec. All three sync into Agents Schema, an open, warehouse-native metadata standard. Claude and Snowflake Cortex read the same context through different doors.

Crahen said the payoff is that fixing a wrong answer means merging a change in dbt rather than editing an opaque agent config. "The result is the 100th question gets the same quality answer as the first," she said. Next steps include automated regression tests, extending context beyond one domain, and feeding corrections back into context docs.

Source: dbt Labs Blog