In-depth Interview: The Second Power Shift – When Agents Take Over Database Action Rights

The primary user of databases is shifting from humans to agents. In 2025, Databricks acquired Neon for approximately $1 billion; Neon disclosed that about 80% of its database instances are created by agents. Yang Xinjun, head of Alibaba Cloud Database, noted that in recent months, the number of Alibaba Cloud PostgreSQL instances created by agents has reached three times the cumulative total of the past five years. These numbers are concentrated on PostgreSQL because large model corpora contain the richest documentation for Postgres, making it the most 'understood' by models.

Traditional application access paths are hardcoded during development, but agents receive goals rather than steps: at runtime, they observe the environment, understand the schema, formulate plans, and continuously adjust based on results. Wang Yuan summarized this as 'execution patterns that change in seconds or even sub-seconds.' Yang Xinjun emphasized that past changes targeted data types and scenarios; this time, the fundamental difference is that the user has changed from human to agent.

When agents become direct operators, the traditional development chain is greatly shortened: a programming agent can apply for a database, design table structures, and generate migration scripts on its own, with users unaware of where the database is located. At one insurance company, business personnel can now make claims in natural language, with agents completing the subsequent processes. They don't write SQL but have become users of database services.

Wang Yuan describes the future database form: users open a browser, express their needs in a dialog box, and receive data services that previously only professionals could access. Yang Xinjun adds that agents don't need graphical interfaces; APIs, command lines, and CLIs suffice. Interaction expands from structured queries to text, images, audio, and video, with output shifting from tables to conclusions. Wang Yuan summarizes this as moving from 'SQL in, Table out' to 'Token in, Token out.'

Capabilities are becoming agentified: previously scattered functions like monitoring, governance, and analysis are consolidated into agents. For example, the DAS Agent handles change recovery, while the Meta Agent continuously understands fields, supplements business semantics, and accumulates knowledge, making subsequent queries more accurate. Thus, databases not only store business data but also 'understanding of data' and 'experience in using data.'

The kernel must be reworked for new workloads: agents run 24/7, operate far faster than humans, explore in parallel toward goals, and iterate through trial and error. Key capabilities include copy-on-write-based second-level branching, fast backup and recovery, operation tracking, and permission and load isolation. In terms of cost structure, a single user query may spawn multiple agents and temporary databases; Alibaba Cloud is exploring multi-tenant sharing to spread costs. At the same time, a deterministic base must be retained: critical data like transactions and sales must be real-time and accurate, not reliant on probabilistic models.

As agents enter the operations system, a 24/7 online 'high-level operations entity' is taking shape. The first tasks to be replaced are those with clear rules, high frequency, and verifiable results: monitoring, slow query analysis, statistics collection, standardized changes, and recovery. Wang Yuan admits that 'anxiety is normal,' but the proliferation of information has not diminished the importance of judgment.

Yang Xinjun points out that critical business data such as transactions, sales, and reconciliation must be 'real-time and accurate' and cannot rely on approximate results. Agents can propose technically feasible solutions but may not understand business implications or bear risks. The DBA role shifts from 'the person who executes operations' to 'the person who decides which operations should occur and what the boundaries of machine capabilities are.'

Future DBA capabilities will focus on three areas: collaborating with agents (process restructuring, task package decision-making), understanding industry and business semantics (identifying business meaning behind technical anomalies), and organizing multi-agent systems (defining responsibility boundaries, data access permissions, and conditions for automation interruption). Evaluation criteria will shift from 'how many incidents were resolved' to 'automation coverage, recommendation accuracy, and the scope of high-risk operation restrictions.'

Yang Xinjun believes the window is changing very quickly, and Chinese databases have an opportunity to move from followers to leaders, but they must experiment in multiple directions simultaneously. '90% may fail, but if you catch that 10%, you can build a true leading advantage.' In 2025, Databricks acquired Neon and Snowflake acquired Crunchy Data; everyone is vying for the same position: the data foundation for the agent era.

The first level is being callable: natural language queries, vector search, agent memory, database branching, BaaS, etc. But this level is hard to build a long-term moat—vector search can be added by anyone, protocols like MCP are inherently standardized, and branching and natural language queries will become default features. Supporting agent calls will eventually be like supporting JDBC—just an entry ticket.

The second level is the real differentiator: whether enterprises dare to let agents take action. Allowing report queries versus allowing modifications to customer status are completely different trust levels. Wang Yuan summarizes enterprise requirements as understandable, traceable, accurate, and secure, plus permission isolation and load isolation. Procurement criteria will expand from QPS to the resources consumed by a single agent task; cost units will expand from instances to tasks, models, and tokens. Yang Xinjun compares tokens to future water, electricity, and oil—consumption must be budgetable and limitable.