Superpower Daily

Superpower Daily

Your reading space

Sign in
Google Supports Three Venice XR Projects With AI Interaction an...GSA Sets October Shift to Zero-Fee, Usage-Based OpenAI PricingAnthropic Says It Removed Claude Accounts Tied to Nine Influenc...Pentagon Negotiates $5B Fluidstack Loan for AI Power and Coolin...TSMC Reports Record August Revenue as AI-Chip Demand Holds Stro...Enflame Shares Rise 206% in Shanghai Debut, Backing Its Next AI...Google Supports Three Venice XR Projects With AI Interaction an...GSA Sets October Shift to Zero-Fee, Usage-Based OpenAI PricingAnthropic Says It Removed Claude Accounts Tied to Nine Influenc...Pentagon Negotiates $5B Fluidstack Loan for AI Power and Coolin...TSMC Reports Record August Revenue as AI-Chip Demand Holds Stro...Enflame Shares Rise 206% in Shanghai Debut, Backing Its Next AI...
Swipe

Oracle Wants AI Agents to Pick Trusted Reports, Not Write SQL

The design can make database answers more repeatable by fixing the query underneath them. But it leaves a consequential judgment with the model: selecting the right report and inputs.

By 3 min read
Oracle Wants AI Agents to Pick Trusted Reports, Not Write SQL
Oracle Wants AI Agents to Pick Trusted Reports, Not Write SQL

Listen to this story

The audio brief

About 1:28
0:001:28
Read transcript
Oracle is proposing a safer way for AI agents to answer questions about production databases: instead of writing SQL on the fly, an agent chooses from reports that analysts have already defined and approved, then supplies typed inputs such as a region or quarter. The reason is that broken SQL is often obvious. The more dangerous failure is a query that runs successfully but applies the wrong business definition—say, treating “month” as days since a customer’s first purchase instead of a calendar month. Oracle cites an August 21 snapshot of the BIRD benchmark, where the best listed model reached 80.04 percent execution accuracy, compared with 92.96 percent for humans. Oracle notes that leaderboard positions can change, and execution accuracy alone does not prove the business logic is right. With fixed SQL underneath each report, analysts become the main validation point. The model’s remaining judgment is consequential: selecting the right report and parameters. Oracle says many teams will still use text-to-SQL, and it is still learning where this governed approach breaks. Access controls limit the blast radius. Agents can inherit a user’s corporate identity, while Deep Data Security enforces row-, column-, and cell-level policies. Oracle also warns that MCP implementations have faced more than 40 reported vulnerabilities this year, amid an NSA warning that safeguards are lagging. The key constraint is simple: repeatable queries do not remove the need to choose the right report.

Story brief

3 key points

Oracle is proposing a governed alternative to live text-to-SQL: analysts predefine reports, while an agent selects among them and supplies typed parameters. The approach targets business-logic errors that produce plausible but incorrect answers, though report selection can still fail and many teams will retain text-to-SQL. An August 21 BIRD snapshot put the best model at 80.04% execution accuracy versus 92.96% for...

  1. 01

    Fixed SQL makes analyst validation and report definitions the main control point, but does not guarantee the agent chose the right report.

  2. 02

    Oracle cites BIRD execution accuracy of 80.04% for the best listed model versus 92.96% for humans; leaderboard positions may change.

  3. 03

    Agents can inherit a user’s corporate permissions, with Deep Data Security enforcing row-, column- and cell-level policies.

AI agents promise plain-English answers from production data. Oracle’s design starts by limiting that promise: rather than let a model generate SQL live, it exposes pre-approved analyst reports as governed Model Context Protocol tools. The agent selects a report and supplies typed inputs such as a region or quarter.

The costly failure is a plausible wrong answer

Oracle is targeting a failure more dangerous than broken syntax: SQL that runs but applies the wrong business definition or table relationship. Its example is interpreting “month” as days since a customer’s first purchase rather than a calendar month. The result can look authoritative while measuring the wrong thing.

Oracle also cites an August 21 BIRD benchmark snapshot: the best listed model reached 80.04% execution accuracy, versus 92.96% for humans. The company notes that leaderboard positions can change; execution accuracy also does not settle whether a query used the right business logic.

Repeatability shifts the model’s job

The proposal makes analysts responsible for defining and validating the query in advance. With a fixed query, Oracle says the same question follows the same trusted path; a changed result should come from changed data, rather than different SQL generated by the model.

That does not make every answer correct. It moves a key model decision to choosing the relevant report and parameters. Oracle says relatively few organizations work this way today, it is still learning where the approach breaks, and many teams will continue using natural-language-to-SQL.

Permissions determine the blast radius

Oracle says an agent can authenticate through an existing corporate identity system and inherit the permissions of the person it represents, rather than receive a set of its own. Its Deep Data Security product applies SQL-declared policies at the row, column and cell levels, including for agent-mediated access.

Those controls do not eliminate prompt injection, Oracle says. They limit how much data an agent can reach if something goes wrong. Oracle cites a May NSA publication warning that MCP adoption had outpaced safeguards, and says more than 40 vulnerabilities had been filed against MCP implementations this year.

Memory and deployment remain part of the package

Oracle AI Agent Memory adds working memory, long-term factual memory and procedural memory to Oracle AI Database. Oracle argues that when agent memory contains company information, it should sit under the same access rules, backups and auditing as other enterprise data.

Four connection options Oracle describes

  • A local option for a developer machine, and a connection built into Autonomous Database.
  • A managed Oracle Cloud Infrastructure service that can reach Oracle databases in Amazon, Microsoft and Google clouds.
  • A self-managed option for data that remains in a customer data center.

Sources

  1. blogs.oracle.comWhat we learned about letting agents into a production database

Loading discussion...

Related coverage

Continue reading