Superpower Daily
Trump Administration Backs OpenAI’s Fair-Use Defense in New Yor...Anthropic’s Fable 5.1 Takes Benchmark Lead, but Its Top Setting...Google Launches Fairwind to Give Selected Defenders AI Tools fo...Loudoun County Confronts Data-Center Backlash After Reaping a T...AWS Makes Its Contact-Center AI Designer Generally Available Wi...Enflame’s Retail IPO Tranche Is Subscribed 4,073 TimesTrump Administration Backs OpenAI’s Fair-Use Defense in New Yor...Anthropic’s Fable 5.1 Takes Benchmark Lead, but Its Top Setting...Google Launches Fairwind to Give Selected Defenders AI Tools fo...Loudoun County Confronts Data-Center Backlash After Reaping a T...AWS Makes Its Contact-Center AI Designer Generally Available Wi...Enflame’s Retail IPO Tranche Is Subscribed 4,073 Times
Toolspublished

Oracle Publishes AI Agent Blueprint That Keeps Database Writes Under Human Approval

Oracle’s supply-chain sample connects chat interfaces to inventory recommendations, but limits execution to a procedure-backed, audited database path after a planner approves it.

By 2 min read
Oracle Publishes AI Agent Blueprint That Keeps Database Writes Under Human Approval
Oracle Publishes AI Agent Blueprint That Keeps Database Writes Under Human Approval
Sign in to save and react

Listen to this story

The audio brief

About 1:37
0:001:37
Read transcript
Oracle has published a supply-chain reference application where an AI agent can recommend an inventory transfer, but cannot execute it on its own. A planner has to approve the transfer first. Only then does Oracle AI Database run the procedure, check feasibility and authorization, lock the inventory, complete the transaction, and create an audit record. The design is deliberately narrow. Oracle’s Java MCP Toolkit exposes just five named operations: find recommendations, retrieve their details, reserve a transfer ID, approve the transfer, and count completed transfers. Generic read and write queries, table access, and administration tools are disabled. The agent gets a specific supply-chain-exchange allowlist, rather than a direct route into the database. The sample also separates the user interface from the governed action path. With A2UI, the agent describes controls in declarative JSON, and the host validates and renders approved native components. MCP Apps use a different model: developer-built HTML and JavaScript delivered as sandboxed UI resources. They are two presentation approaches over the same controlled service, not interchangeable features. MCP Apps are demonstrated in ChatGPT, Claude, and Gemini Enterprise. A2UI support depends on the host and protocol: Gemini Enterprise uses version zero point eight over A2A, while the standalone browser uses version zero point nine point one through AG-UI events. The constraint to watch is compatibility: each host must validate the version and component catalog it supports, so this is a blueprint for bounded actions—not universal agent access.

Story brief

3 key points

Oracle’s reference application demonstrates a tightly bounded pattern for database-connected agents: the model can discover and prepare an inventory transfer, but a planner must approve it before execution. The Java MCP Toolkit exposes five purpose-built operations instead of general database tools, while Oracle AI Database retains authorization, transaction, feasibility, locking, and audit responsibilities. The UI...

  1. 01

    The Toolkit exposes five operations: find recommendations, retrieve details, reserve an ID, approve transfers, and count completed transfers.

  2. 02

    Oracle disables generic read, write, table, and administration tools, limiting the agent to the supply-chain-exchange allowlist.

  3. 03

    A2UI uses host-validated declarative components; MCP Apps deliver developer-built HTML and JavaScript in sandboxed ui:// resources.

Oracle has published a supply-chain reference application for showing database-calculated inventory-transfer recommendations in AI chat interfaces. The apparent autonomy stops at the approval line: a planner must approve a transfer before Oracle AI Database executes the procedure, validates and locks inventory, and creates an audit record.

Two interface paths, not one generated frontend

The sample offers two ways to present the task. A2UI lets an agent describe controls in declarative JSON, while the host validates that request and renders only components in its approved native catalog. That lets a task-specific interface appear without running agent-generated frontend code.

MCP Apps take a different route: they package developer-built HTML and JavaScript as sandboxed ui:// resources for compatible conversational hosts. The two approaches are separate UI models over the same governed service; using one does not automatically enable the other.

The tool list replaces direct database access

Oracle configures the Database MCP Java Toolkit around those operations rather than giving the agent general database controls. Built-in read-query, write-query, table, and administration tools are not enabled, and the application requires the supply-chain-exchange allowlist. Oracle AI Database remains responsible for feasibility calculations, authorization, transactions, and auditing.

Host support depends on the adapter

The MCP App route is shown in ChatGPT, Claude, and Gemini Enterprise. Gemini Enterprise also receives A2UI v0.8 DataParts over A2A, an agent-to-agent communication protocol. The standalone browser instead receives A2UI v0.9.1 through AG-UI events.

Those versioned paths are alternatives, not a promise of universal compatibility. Each host must validate the A2UI version and component catalog it advertises, and Oracle says product support should be verified rather than assumed. The sample is therefore a concrete architecture for bounded agent actions, not an off-the-shelf autonomous inventory system.

Sources

  1. blogs.oracle.comDevelop A2UI and MCP Apps with Oracle AI Database and the Java MCP Toolkit running in Google Gemini Enterprise App, ChatGPT, and Claude

Related coverage

Continue reading