Session Overview
Discover how OpenAI transformed Apache Flink development by embedding end-to-end streaming workflows directly into Codex.
In this talk, Stella (Technical Staff at OpenAI) walks through how internal engineers build, debug, and deploy production-ready Flink pipelines using natural language. Learn how OpenAI uses custom Codex plugins, specialized skills, and Model Context Protocol (MCP) tools to package platform expertise, enforce production readiness, and slash debugging times from hours to minutes.
What You'll Learn
- Why OpenAI turned Codex into the default interface for Flink engineering
- The architecture behind the AI agent runtime: Context, evidence, and tooling
- How routing skills keep prompt context small and avoid token bloat
- Catching dangerous job graph state changes and checkpoint incompatibilities before deploy
- Symptom-based debugging for lag, bottlenecks, and TaskManager exceptions
- Safeguarding production with human-in-the-loop approvals for destructive actions
- Key metrics: 80%+ agent-driven operations and a 90% drop in support requests
Chapters
- 00:00 — Introduction & the AI-native Flink vision
- 02:12 — Traditional Flink workflows vs. the Codex experience
- 05:19 — Interactive learning, architecture visualizer & job graphs
- 07:51 — Production readiness checks & guardrails
- 08:48 — Agent architecture: Context, evidence, and MCP tooling
- 10:53 — Efficient context routing with specialized skills
- 14:18 — Preventing incompatible state & checkpoint failures
- 15:46 — Symptom-based debugging & bottleneck diagnosis
- 17:48 — Dedicated Flink MCP server design & safety annotations
- 19:41 — Key benefits, internal impact & wrap-up
