How Meta Built Forum With AI-Native Engineering

This post details the application of agentic development to build the Forum app, focusing on a 'human-on-the-loop' model where AI agents manage iterations (plan, implement, build, test, correct) under human supervision. Key advancements include rigorous upstream decomposition for planning, integrated automated verification (build status, unit tests, UI spec matching) within the agent loop, simulated user testing with navigation agents, and agentic production operations for root-cause analysis. The operating model involves rules, skills, agents, commands, and hooks. Architectural changes prioritized legibility and modularity for agentic throughput, and agents were given tools (e.g., live view hierarchy checks) rather than just descriptions. A separate build flavor sandbox was used for safe AI prototyping. AI stress testing with agents identified and fixed a memory issue during development. The approach was applied to building the stack (architectural mapping) and operating the stack (agent-driven investigation of user feedback).