Governed AI agents powered by Oracle AI Database
Today we are pleased to announce the general availability of Oracle AI Database Agent Engine in Oracle AI Database 26ai. Agent Engine is the framework for defining, executing, remembering, and governing agents. At its core is an in database agent execution runtime that runs declarative workflows with governance and transactional consistency. It is designed for workflows that are data centric, latency critical, and operate on confidential, business critical data, and it separates the execution of a workflow from the content that is disclosed outside the database.
Key highlights:
- Agent Engine runs AI agent workflows inside the database; alongside the data.
- Using agents built with Open Agent Specification, Agent Engine supports a rich set of workflow patterns, such as ReAct, Sequential, Parallel, Orchestrator, Router, and Loop.
- The database provides the execution environment for the workflow’s tools, maintains its state, and handles retrieval. Language models are invoked only at the points defined in the workflow, eliminating the need to build, secure, and synchronize a separate orchestration tier with your data.
Declarative workflows with governed execution
Agent Engine’s in-database agent execution runtime runs declarative AI agent workflows inside the database, close to the data they use, enabling transactionally consistent execution of in-database operations and workflow states. Workflows can include sequential, parallel, ReAct, orchestrator, router, and loop patterns, as well as human approval gates. SQL and PL/SQL handle deterministic operations such as joins, calculations, and rule evaluation, while LLMs are invoked only when interpretation is needed.
Workflow state and sensitive intermediate results are protected by database policy controls, resource limits, and auditing. Model calls receive structured inputs scoped to each step, not database connections or unrestricted table access, and can follow only approved paths. Because the framework is based on the Open Agent Specification, workflows are portable across compatible agent implementations.
Governed information flow
Agent workflows often need privileged information to calculate a result, but the model or calling application does not always need the underlying values. Oracle AI Database keeps privileged computation inside trusted database operations and passes only the permitted result to later workflow. Furthermore, for sensitive data, the decision contract for the operation’s output can prepare an approved contract to pass the minimal information needed for LLM to reason about or write a response.
Human-in-the-loop and traceability
Human approval is a first-class workflow step. A workflow can run autonomously and pause before any consequential action, routing the request through existing application controls. Every tool call, model call, and decision is recorded along with its inputs and sources, giving reviewers and auditors a complete account of each run.
Data-centric workloads
Workflows that involve multiple database operations or generate large intermediate results can benefit from using the in-database agent executor runtime; the runtime executes the workflows without an additional orchestration service dedicated to them. Passing the output of one operation directly to the next within the database reduces latency. It also keeps intermediate data close to its source, avoiding repeated network round trips and the associated data transfer and orchestration overhead of an external service. Additionally, workflows that operate on sensitive, confidential business data can avoid oversharing information by executing the agents and workflows inside the database boundary and setting up proper disclosure contracts.
Database-grade operations
A workflow running on Agent Engine inherits the mission critical capabilities of Oracle AI Database. Workflow state is stored in the database alongside the data used by the workflow, keeping database operations and their state transactionally consistent. Because execution remains within the database, resource management can isolate agent workloads from production workloads; and the database’s encryption, access-control, and auditing capabilities apply throughout the workflow. Deep Data Security further carries end-user identity through data access and enforces the relevant policies at the source. If a model or retrieval dependency becomes temporarily unavailable, the workflow can pauses at its last committed state and resume when the dependency recovers.
Getting started
Ready to build? Visit the Oracle AI Database Agent Engine product page for documentation, samples, and tutorials. For a complete worked example, an agent that screens invoices against service evidence with a human decision gate read Catch Revenue and Service Exceptions Before They Become Customer Problems.