Postgres Summit US2026

PostgreSQL is usually treated as a passive datastore in AI systems while orchestration logic and workflow control live elsewhere in the stack. This talk presents an architecture where PostgreSQL becomes the control plane that governs AI workflows.

Large language models are inherently probabilistic, which makes them difficult to operate reliably in production environments. By placing a deterministic system around the model and using PostgreSQL to manage workflow state, validation, verification, and orchestration, we can constrain this stochastic core and make the overall system far more predictable, observable, and accountable.

This talk introduces a practical design framework for building AI workflows around PostgreSQL. We will examine schema patterns for storing workflow state, validation and quality gates that prevent incorrect outputs from propagating through a pipeline, and verification and audit patterns, built from constraints, transactions, roles, triggers, and append-only ledgers, that ensure a step cannot be marked complete without evidence and that the component doing the work cannot rewrite the verdict on its own output.

Using practical examples drawn from production systems, the talk shows how PostgreSQL can orchestrate Retrieval Augmented Generation pipelines and MCP-driven agent workflows, and, more importantly, how its primitives make those applications production ready: verifiable, reliable, correct, and resistant to reward-hacking, by enforcing deterministic control over probabilistic model behavior.

The talk closes with a reference schema and framework that attendees can adopt in their own pipelines to build reliable agentic systems that lower the risk of running such software in production and at scale.