Andy Pavlo: What AI Agents Change for Databases
What happens when agents start creating databases, reading data, and modifying it around the clock? That's the subject of Matt Turck's interview with Andy Pavlo on October 8, 2026. Pavlo researches databases at Carnegie Mellon and has also been leading ClickHouse Labs since August.
Several topics stood out:
🔸 Workloads from agents
Pavlo predicts that query volume could rise 10–100 times. He doesn't yet know how much query complexity will change. It's a possible future scenario, rather than market-wide growth that has already been measured.
🔸 Permissions and isolation
Agents need restricted permissions and separate database branches for experiments. Pavlo thinks familiar data protection mechanisms still apply here. Manually approving every action by a hundred agents won't scale.
🔸 Agent memory
Behind the “files or databases” debate, he suggests looking at the guarantees for storing and updating state. SQLite comes up as a simple option.
🔸 Text-to-SQL
Pavlo connects quality to the semantic layer and context. The agent needs to know how to turn a question into the right query.
🔸 Automatic tuning
In his rough estimate, an LLM can get 80–85% of the way to a good configuration for certain parameters in 15 minutes. Hard cases still require specialized methods. The interview doesn't explain how he arrived at that estimate.
🔸 Convergence of OLTP and OLAP
Pavlo looks at why analytics platforms need PostgreSQL and how separate services that move data differ from two engines operating on one copy. A shared interface tells you little about the internal architecture.
We've covered one of Pavlo's papers on Research Insights Made Simple before: “What Goes Around Comes Around… And Around…,” written with Michael Stonebraker. In August 2025, Alexey Svetlichny and I devoted two episodes to it: data models and query languages and systems architecture. It's a good overview of about twenty years of database development. The authors make two related points. SQL and the relational model keep absorbing ideas from their competitors, while the engines themselves change substantially with hardware and workloads. Columnar storage, the cloud, and the separation of storage from compute can make the system behind familiar SQL look quite different.
On October 12, 2026, Nikolay Golov and I also discussed modern data platforms and four ways to bring data together for AI: copies and data marts, federated queries, a general-purpose DBMS, or a shared storage layer. Pavlo's discussion of combining transactions and analytics complements our map. The same label can hide very different trade-offs in freshness, isolation, and cost. I wrote a detailed long read about it too. It goes beyond choosing an engine: who defines the meaning of metrics, whose permissions the agent uses to read the data, and how we verify an action taken on the basis of an analysis. Identifying customers who should get a discount and actually granting it are two different engineering tasks.
You can start with the interview, then use our discussions to explore how the platforms work and where these trade-offs came from.
#Data #AI #Database #Architecture #ResearchInsights