Oracle introduces an AI database query agent grounded in business rules

Oracle is giving its private-agent platform a way to check what a business question means before turning it into a database query. Announced October 26, 2026, Private Agent Factory 26.10 introduces Deep NL2SQL, an agent designed to combine business definitions, user clarification and tested knowledge changes rather than rely on database structure alone.

Teaching the agent what a sale means

SQL is the language used to query database tables. A statement can execute successfully while applying the wrong business definition. Oracle illustrates this with top-selling products: sales could mean units, revenue or profit, and might count only after the return window closes.

Deep NL2SQL addresses that gap through a Business Knowledge Graph, an organized collection of database relationships and business context. Application experts use a Knowledge Workbench to group tables, metric definitions, rules, documents and example questions into topics. Oracle says search selects relevant topics instead of presenting the model with the entire enterprise database structure.

If that context leaves the question ambiguous, the agent can ask the user to clarify. It plans, generates and executes SQL, then reviews its answer. Users can inspect the generated SQL rather than receive only a final answer.

Corrections pass through experts and tests

The agent logs conversation feedback, including votes on results. Experts review feedback and curate useful knowledge rather than automatically adding every user response to the shared knowledge base.

Changes also have a staging step. Teams register representative questions and validated SQL as a “gold” test set. Oracle’s Deep NL2SQL technical explanation describes how proposed knowledge edits remain unpublished while an evaluation compares generated answers with expected results, checking for disruption of previously validated behavior.

Connecting and inspecting private agents

The same release adds a Python software development kit and standardized REST interfaces for connecting agents to applications. Expanded Agent-to-Agent support handles structured content and tasks that need not finish immediately. An Ingestion Service makes prepared knowledge reusable across research and custom workflows.

Native observability lets authorized users inspect execution traces, latency, token usage and errors. Teams can distinguish time spent in model calls, tools and other workflow steps when investigating slow agents or unexpected results.