Schedule - PGConf NYC 2025

Beyond monitoring: How we built an open-source, self-healing Postgres agent

Date: 2025-10-01
Time: 12:00–12:50
Room: Hub One
Level: Intermediate
Feedback: Leave feedback

We’ve moved from buzzing pagers to polished SRE dashboards, yet incident response is still largely reactive: collect signals, page a human, hope they’re sharp at 3 a.m. In this talk, I’ll explore what happens when we flip that script. Shifting from reactive alerts, to proactive self-healing in the form of an open-source agent for PostgreSQL.

• Signals → Playbooks. pg_stat_statements, cloud metadata and more flow into an LLM-driven engine that selects an executable playbook. Each playbook defines diagnostics, safe-change limits, rollback steps, and can be tuned to your business needs.
• Pluggable tooling. Need git history, Sentry error context, a read-only peek at another Postgres cluster, or live Datadog metrics? Plug in the matching MCP server—or one you build yourself—and the agent can call it instantly.
• Customizable evals. Built-in evaluation hooks spin up a shadow workload or staging clone so every CREATE INDEX, VACUUM, or config tweak proves its worth before rolling out to production.
• Trust & approval. The agent can run arbitrary SQL, but approval gates let you decide which actions are fully automated and which still need a human click.

As the head of product at Xata, I’ll highlight real-world examples across Xata’s fleet of serverless Postgres instances and map out where we think agents like this will go next. Bring your skepticism and leave with a PostgreSQL agent you can POC on Monday.

Slides

The following slides have been made available for this session:

• Beyond monitoring_ How we built an open-source, self-healing Postgres agent.pdf

Speaker

Alex Francoeur