Oracle connects LangChain’s AI agents to its database and cloud models

Developers can now connect LangChain’s task-planning agents to Oracle’s cloud models and database through a native integration. Oracle’s November 10 announcement adds search, saved work and human-approval hooks to its open-source langchain-oci package. Its accompanying clinic demo illustrates a separate boundary: agents prepare recommendations, while database rules and people control what happens next.

The new entry point, create_deepagents_agent, connects LangChain Deep Agents to OCI Generative AI and Oracle AI Database 26ai. Deep Agents is LangChain’s open-source framework, built on LangGraph, for work that involves more than a question and an answer. It adds a plan, specialist agents with their own tools and working context, and files for keeping notes and drafts outside the model’s immediate conversation.

Oracle’s integration generates tools for inspecting store statistics, searching records and retrieving documents. Search combines vector matching with keyword matching and routes requests across the configured stores. Oracle AI Database 26ai and OpenSearch indices are supported. Developers can also supply their own SQL tools, using the language for querying database tables.

Saved work has three parts: OracleSaver records checkpoints for a running task, OracleStore holds long-term memory, and StoreBackend preserves files. Those database-backed records let a run be inspected, resumed or continued days later. The framework also supports structured output, reusable skills and approval before selected tool calls. Model access is not limited to Oracle’s service: developers can bring another LangChain chat model that supports tool calls.

The accompanying demonstration shows how those building blocks fit a controlled process. A team prepares a clinic visit by reading a chart, checking guidelines and research, proposing tests and follow-up, and flagging a risky medication. The resulting brief points back to database records. All patients in the example are invented.

Reading and writing follow different paths. The chart analyst can generate SQL queries, but the database applies row-level security—a filter that restricts which records a database user can access. Each patient’s agent uses its own database account. A request for all lab results therefore returns only the assigned patient’s records, regardless of the query the model writes.

Writing is narrower. An agent fills in typed fields describing a proposal, and tool code saves them through a fixed database statement. The model does not write that statement. A database trigger, a rule applied when a record is added, rejects medication-change proposals unless they come from the medication-safety agent. Tool code supplies the agent’s identity; the model does not set it.

The package is available on PyPI with the deepagents extra. The full framework requires Python 3.11 through 3.13; a lightweight agent loop supports Python 3.9 and later. Oracle supplies an integration guide and runnable tutorial in its repository.

The demo also places agents inside an NVIDIA OpenShell sandbox limited to reaching the database and OCI Generative AI. A separate program checks cited facts before the brief is read. Oracle reports 41 successful end-to-end checks, alongside 25 database-rule checks and seven sandbox probes. These are company-reported tests of the demonstration, not results from treating patients.