Roblox Engineering Director: "Prompt to Prod" Is Broken Without Trust
AI can generate code in seconds, but can your infrastructure handle autonomous deployments without causing a major SEV? Andrew Swerdlow, Engineering Director at Roblox (ex-Google, ex-Instagram), breaks down how Roblox transitioned from basic AI autocomplete to fully autonomous "Prompt to Prod" pipelines safely. Discover how to solve the "typing vs. trust" paradox by implementing automated code review agents, sandboxed runtime environments, institutional knowledge extraction ("exemplars"), and new metrics to track engineering productivity in an agentic AI world. ⏱️ Video Timestamps (For Navigation) 00:00 - The AI Paradox: We Solved Typing, Not Trust 01:12 - Why Autonomous Code Generation Creates Technical Debt 02:15 - Roblox's 3 Pillars: Alignment, Security, and Productivity 03:08 - Security Foundations: Sandboxes & Policy Gateways 04:22 - Mining 700k PRs for "Exemplars" to Train AI Reviewers 07:15 - Achieving a 70% Acceptance Rate on AI Code Reviews 08:05 - The "Prompt to Prod" Journey: Infrastructure & Plumbing 10:48 - Rethinking Engineering Productivity Beyond Lines of Code 13:10 - Measuring Agent Quality and Long-Running Turn Rates 15:42 - Q&A: Addressing SEVs, Dissecting Design Docs, and the Future of Engineering 🔗 Transcript & slides available on InfoQ: https://bit.ly/45QbQUR #SoftwareEngineering #AIAgents #DevOps #Roblox Engineering