PGConf.PL2026

Databases were designed for humans: long-lived connections, predictable traffic patterns , schemas that evolve slowly and operators who provision, configure and monitor infrastructure. Agent-driven systems are beginning to break those assumptions. They generate bursty workloads at a scale and unpredictability that traditional database design never anticipated. If PostgreSQL is going to remain the default database in an agent-driven world, the ecosystem needs to rethink what "database infrastructure" means at a fundamental level.

This talk examines what it actually means to build an AI-native database platform. Not "has vectors." AI-native means the platform is optimized end-to-end for agent workflows, which look fundamentally different from human application patterns.

We'll walk through four emerging agent-database interaction patterns. Tool-using agents (the most common today) never connect to a database directly; they call tools that wrap database operations through MCP servers, issuing many small, short-lived queries with strict safety constraints. Evaluation and simulation environments run thousands of isolated agent executions against the same dataset for regression testing, model comparison, and training data generation; requiring fast cloning, strong isolation and low cost per environment. Autonomous development agents spin up environments, modify schemas, replay traffic and validate changes, treating the database as disposable infrastructure rather than persistent state. And agents increasingly use databases as hybrid memory stores, combining relational data, vector search and document storage for conversation history, structured facts and intermediate reasoning.

These workloads demand infrastructure primitives that traditional database platforms were not designed around. At the same time, AI-native infrastructure cannot come at the expense of reliability. Platform availability, transactional correctness, durability and operational safety remain non-negotiable. The goal of the talk is not to predict the future of AI agents, but to examine how agent-driven development is already changing the assumptions databases and data platforms were originally built around and what PostgreSQL needs to remain the default platform in that world.