Deep NL2SQL and new ways to connect and operate private agents

Business users want answers they can act on. Developers want to bring agents into the applications people already use. Platform teams need to understand how those agents behave and keep control of the data and govern th systems they access.

Oracle AI Database Private Agent Factory 26.10 addresses these needs.. This release introduces Deep NL2SQL, a major addition for asking business questions about enterprise data, alongside new Python SDK and REST integration capabilities, reusable ingestion services, expanded agent interoperability, and native observability. 

Private Agent Factory gives teams a visual environment to build, test, and run private agents with Oracle Database, enterprise documents, APIs, and supported models. This release extends that foundation from building an agent to connecting it to a business workflow and understanding its operation.

Deep NL2SQL brings business knowledge into every question

A database schema tells a model where data is stored. It rarely explains what the business means by an open invoice, a completed order, or a top-selling product. Without those definitions, a SQL statement can run successfully and still answer the wrong question.

Deep NL2SQL addresses this gap with a Business Knowledge Graph. Teams organize relevant tables, business definitions, metrics, rules, documents, and example questions and SQL into topics. The agent finds the topics relevant to a question and uses that context to plan and generate SQL. When a question is ambiguous, it can ask the user for clarification before continuing.

Financial Control and Business Operations v2 with 350 tables and views.
Supply Planning and Fulfillment v2 with 386 tables and views.

Figure 1. Knowledge graphs for finance (350 tables/views) and supply planning and fulfillment (386 tables/views).

Turn expert knowledge into a resource the whole team can use

Consider a finance team asking which supplier invoices need attention. The answer depends on the definition of outstanding balance, due dates, payment status, currencies, and organizational scope. Deep NL2SQL gives application experts a place called the Knowledge Workbench to capture those conventions, so business users do not have to reconstruct them in every conversation. 

The Knowledge Workbench supports the creation and maintenance of business context. Its NL/SQL Collection brings natural-language questions and corresponding SQL together for inspection and reuse. The agent makes generated SQL available for review, helping users understand how an answer was produced. 

Natural-language questions and their SQL statements in the EBS knowledge workspace.

Figure 2. The NL/SQL Collection presents questions and SQL side by side so experts can inspect and maintain examples.

The release also brings Knowledge, Memory, and Annotation Workbenches into the workflow for maintaining semantic context. Experts can review feedback and curate useful knowledge, rather than treating every user response as an automatic change to the shared knowledge base.

An evaluation workflow lets teams compare proposed knowledge changes with representative questions and reference SQL before publishing those changes. Together, clarification, SQL review, evaluation, and curated feedback help teams improve their agents as business definitions and data evolve.

Connect private agents to the applications people use

The new Agent Factory Python SDK gives developers a supported way to run Custom Workflows, Knowledge Agents, Data Analysis Agents, and Deep Data Research Agents from Python applications. Supported authentication options include username and password and browser-assisted SSO, while operations continue to follow the user’s existing Agent Factory permissions.

The SDK and standardized REST APIs provide integration options for applications, scripts, and business workflows. The SDK also supports operations on supported model connections and data sources. Teams can bring agent capabilities into an existing application without asking every user to work directly in the Agent Factory interface.

Connect agents across systems

The expanded Agent-to-Agent capabilities support additional content types, including JSON and Markdown, and asynchronous task handling. These improvements help an Agent Factory workflow to exchange structured information with a compatible external agent and handle work that takes longer than a single immediate response.

For customers and partners building solutions across applications, the SDK, REST APIs, and A2A support provide several ways to connect agents to an existing architecture.

Prepare knowledge once and reuse it across workflows

The new Ingestion Service makes knowledge preparation a reusable part of the platform. Teams can configure supported sources and ingestion settings, then use the prepared knowledge in research and custom agent workflows. The Deep Data Research Agent integration brings ingestion and retrieval into a more consistent workflow, with improvements to source handling and chunking.

Knowledge Agent also gains OpenRouter support, extending the available model choices. Updated Oracle AI Agent Memory integration and platform improvements support the broader work of building and maintaining agents over time.

Understand agent behavior with native observability

When an agent is slow or returns an unexpected result, teams need more than the final answer. Native observability brings metrics, sessions, traces, and spans into Private Agent Factory so authorized users can investigate agent behavior in one place.

Teams can examine execution activity, latency, token usage, and errors, then move from an overall view to a particular session or trace. This helps distinguish time spent in model calls, tools, and other workflow steps, and gives developers a practical starting point for troubleshooting and improvement.

Start with a business workflow

For business teams, Deep NL2SQL offers a way to ask questions using shared business knowledge and review the SQL behind an answer. For developers and partners, the release adds integration options for bringing agents into applications. For platform teams, reusable ingestion and native observability make it easier to prepare knowledge and understand agent execution.

Choose a focused business question, build the knowledge and examples it needs, evaluate the results, and connect the agent to the workflow where people will use it.