Enterprise Agents Are Moving Into the Data Layer, Not On Top of It

Rules enforced where the data lives

Oracle announced the Deep Trust AI architecture, which targets four layers: software platform, data, data access, and application development, the company says. The design puts security and governance into the database layer rather than leaving them to each application that calls a model. m:…5657

Two components carry the idea. Oracle Deep Data Security enables organizations to define end-user data access rules directly in the Oracle AI Database, the company says, and Oracle SQL Assertions enforce business rules directly in the database to prevent incorrect actions or conflicting data. Oracle Trusted Data APIs provide controlled read and write access to logical business objects instead of unrestricted SQL access, the company says. m:…5657

The same announcement adds Oracle Deep NL2SQL, which uses business topics and metadata to improve the accuracy of natural language queries, the company says, with Continuous Feedback Learning for user corrections and Knowledge Change Verification to reverify previously validated answers before publishing changes. The pattern is consistent: business meaning and business rules sit with the data, not in a prompt. m:…5657

Agents without moving the data

Oracle's Autonomous AI Lakehouse update applies the same logic to analytics. The lakehouse allows AI agents to access data across Oracle and non-Oracle databases, Apache Iceberg tables, and object storage, the company says, and users can now create, populate, update, and delete Apache Iceberg tables using Oracle SQL. m:…4812

The AI Data Catalog allows mounting of external catalogs like AWS Glue, Databricks Unity Catalog, and Snowflake Horizon Catalog, the company says, and OpenLineage support is integrated to extend lineage visibility across the data ecosystem. Hierarchical pruning in the catalog narrows search from catalogs to columns to reduce metadata processing costs, the company says. m:…4812

Databricks made a parallel move on the search side. Databricks AI Search now natively supports Streaming Tables, Materialized Views, and Managed tables using Iceberg v3+, the company says, so teams can build search applications directly on live data without an intermediate Delta table synchronization step. Streaming Tables are suitable for continuously arriving events, and Materialized Views enable personalization, session-aware search, and agent assist, the company says. m:…4279

What breaks when the plumbing is wrong

Writer's experience shows the cost of getting the storage layer wrong. Writer routes every customer request through LLM inference providers, and its tracing pipeline generates up to 600 rows per minute at peak, with its primary traces table receiving roughly 1.3 million rows per day. On standard Snowflake tables, insertions failed 45% of the time due to lock contention. m:…7424

After switching to Snowflake Hybrid Tables, which use row-oriented storage with row-level locking, P50 insertion latency dropped from 2,362ms to 154ms and write failures dropped from 45% to near zero. Writer also downsized the warehouse powering its ingest pipeline from medium to extra small, and now uses a Snowflake Task to incrementally copy rows from the hybrid table to a standard table for downstream analytics. m:…7424

The lesson generalizes beyond one pipeline. Agent workloads produce high-velocity operational data — traces, evaluations, tool calls — and a table type designed for analytics batch loads will fail under that write pattern. Choosing the storage shape for the write pattern, rather than for the query pattern alone, is part of agent infrastructure now. m:…7424

What to check before adopting

The vendor claims here are all directional, not measured for a buyer's own workload. Oracle's lakehouse benchmarks, such as Lake Cache reducing remote data access time by up to 8X in benchmark testing, describe Oracle's tests, and Databricks' claim that native support reduces pipeline overhead describes its own architecture, not a customer's measured saving. m:…4812

Teams evaluating these platforms should test three things on their own data: whether permissions set in the database actually constrain what the agent can read, whether search over live tables returns current rows without a sync lag, and whether tracing and logging keep up at peak write rates. Those are the failure modes the announcements and the Writer case each illustrate. m:…5657

The strategic question is simpler than the feature lists. If agents are going to act on enterprise data, the enterprise needs the agent's permissions, business definitions, and audit trail to live with the data — because a permission that only exists in an application prompt disappears the moment another application connects to the same table. m:…5657