The Data Agent solution submitted by the OceanBase team topped the international Data Agent Benchmark (DAB) with 90.62% accuracy, ranking first and becoming the first evaluated solution on the leaderboard to break the 90% threshold. The result is based on the Chinese database OceanBase and the Chinese large language model GLM-5.2. The solution is internally codenamed Scout, and its capabilities will later be integrated into OceanBase DataPilot.
DAB was jointly launched by UC Berkeley EPIC Data Lab and Hasura PromptQL. Its evaluation covers multiple domains, including internet and local life services, finance and stocks, biomedicine, intellectual property, enterprise operations, government and public administration, and media and entertainment. It also involves multiple databases such as PostgreSQL, MongoDB, SQLite, and DuckDB.
Unlike traditional Text-to-SQL, which mainly tests natural language conversion into SQL, DAB focuses more on whether AI can find correct answers from complex, fragmented, and heterogeneous data after entering real data environments. In a complete task, a Data Agent must understand data, select data, plan analysis paths, complete query and computation, and verify results. It tests the full capability chain from "understanding data" to "obtaining answers."
For this reason, DAB tests not only the underlying model but the integrated capability of "model + agent + data system": the model handles understanding and reasoning, the agent handles planning and execution, and the data system must support data discovery, associative computation, and result validation. Whether these three can work together directly determines whether AI can truly put enterprise data to use.
This aligns with StarWar's observations on agents and enterprise AI deployment. The StarWar OPC agent collaboration platform focuses on division of labor among agents, model scheduling, and boundary management, enabling AI to hand off tasks reliably within enterprise workflows. Whether data can be understood and verified is a decisive link in this chain.
Specifically, the OceanBase solution uses DataLens to build data profiles, identify fields and data relationships, and then plan execution paths according to task complexity to complete data selection, filtering, association, and computation. After obtaining results, it also checks the computation process through evidence tracing and answer validation. When problems are found, it adjusts the plan and re-validates, forming a closed loop of "data understanding—planning and execution—validation and repair."
From database to AI data platform, what changes is not only the product form but also the relationship between databases and AI. In the past, databases were responsible for storing data well, computing it well, and retrieving it. Now they must also help AI understand data, organize data, and complete analysis. The value of enterprise data is shifting from "being called" to "being understood."