When Databases' Primary Users Are No Longer Humans: Our Architectural Practices and Reflections in AI Agent Scenarios

Trend observed on TiDB Cloud

It has been reported that TiDB Cloud (TiDB 云) engineers saw a striking shift over the past year: more than 90% of newly created database clusters were not spun up by humans but by autonomous AI agents. Huxiu (虎嗅) republished an InfoQ piece by Huang Dongxu (黄东旭) describing how agents now routinely create databases, generate schemas, run experiments and destroy instances without DBA intervention — often without any human ever knowing. What happens to two decades of database assumptions — capacity planning, schema design, operational playbooks and pricing models — when the primary “user” is a bot?

Architectural responses and trade‑offs

Faced with "one agent, one session, one logical database" at scale, traditional per‑product or per‑tenant models explode. It has been reported that one customer created nearly a million logical tenants in three months, with roughly 99% single‑use. The engineering answers have been multilayered: extreme multitenancy (one physical cluster hosting massive logical tenant counts), aggressive metadata optimizations to survive millions of tables, and radical storage‑compute separation so compute can scale to zero while object storage holds cold data. These come with trade‑offs — cold‑start delays of hundreds of milliseconds, a need for strict resource control to prevent noisy‑neighbour failures, and the uncomfortable truth that AI‑generated SQL looks and behaves very differently from human‑written queries, so standard OLTP benchmarks can be misleading. It has been reported that one migration from MySQL to TiDB required only minor code changes thanks to protocol compatibility, but still demanded hours of query‑plan tuning to match agent query patterns.

Broader implications for builders and operators

Another case, Plaud, highlights the storage challenges of long, multimedia contexts: single context blobs of 30–50MB of text and audio break traditional OLTP assumptions and prompt designs that mix S3‑style object stores, metadata databases and caching. That combination works — but it is brittle, with consistency and tail‑latency pains that drive teams toward more structured, database‑backed approaches when agents need deterministic, auditable analysis. Against a backdrop of global tech tensions and export controls that make hardware efficiency and cost control strategically important, these software architecture choices matter beyond performance: they decide whether an AI agent product is commercially viable. For Western readers building agent‑driven services, the lesson is clear — rethink tenancy, expect unfamiliar query shapes, design for scale‑to‑zero and explicit resource isolation from day one.