Oracle gives Java apps a new way to run database-managed AI agents

Java applications can now manage and run Oracle’s database-defined AI agents through a dedicated programming interface. In an October 23 post, Oracle described the addition in Select AI for Java version 1.1.0. The distinction is where the work happens: Java controls the resources, while the database handles workflow execution, state and permissions.

The update builds on a software development kit that already lets Java applications turn plain-language questions into SQL, the language used to query database tables. Applications can inspect, explain or run the generated query. The SDK also supports answers grounded in retrieved enterprise documents, using Oracle AI Vector Search and Select AI vector indexes.

Java controls; the database executes

The new entry point, SelectAIAgent, exposes agents, tasks, tools and teams as Java resources. Each has a different job: a tool defines an available operation; a task describes work and its instructions; an agent carries a role and configuration; and a team maps agents to tasks and defines how they cooperate.

These objects are not a separate Java agent runtime. The SDK connects through JDBC, Java’s database connection interface, rather than calling AI-provider endpoints directly. Oracle AI Database validates and authorizes requests, manages team workflows and state, invokes configured tools, and handles the required interactions with AI providers.

A workflow can outlast a connection

For applications using pooled database connections, the continuity mechanism is a database-created conversation identifier. Passing the same identifier lets a team workflow continue across separate application requests and JDBC connections. The application must preserve that identifier if it needs to retrieve state, provide human input or resume the run later.

That includes workflows that pause for a person. Oracle’s documentation describes a human tool as a database-defined tool that stops a team workflow for review or required input. Java APIs can retrieve and update supported team state, allowing an application to supply input when a workflow is waiting.

Teams can also be shared across database users through grants. The SDK distinguishes resources owned by the connected user from teams available through direct or role-based access. Its available-team discovery method returns read-only descriptions, rather than mutable team handles. Team definitions can also be imported or exported between environments.

A Java handle is not a deployed agent

Creating a configured Java handle does not create its database resource. Developers must call create() for a new resource, or load an existing resource by name before using its database-backed methods. That separates an application-side object from the database definition whose operations are subject to validation and authorization.

Oracle’s agent-support documentation lists Autonomous AI Database 26ai and 19c as supported environments. That is not a promise that every operation works on every release: the database must support the requested operation and related metadata views. The SDK does not check the database for support for every operation.

Before running a team

• Install the SDK and a compatible Oracle JDBC Thin Driver, then configure database connectivity and authentication.
• Grant the user the required agent privileges, metadata access and permissions for the resources its tools use.
• Configure the necessary AI profiles, credentials, provider access and network access; obtain an owner’s grant for a shared team.

The SDK does not automatically create database users, package grants, credentials, network access rules or shared-team grants. Those remain deployment requirements even though the application can manage agent resources through Java.