Claire Vo's latest episode of How I AI features Alexander Embiricos, OpenAI's product lead for Codex. They focus on how AI coding agents, particularly Codex, improve productivity in software development. Embiricos shares that when half of OpenAI's technical staff used Codex, they produced about 70% more pull requests than non-users. Now, nearly all technical staff consistently rely on Codex, underscoring its effectiveness. Automated code review is a standout feature, implemented across nearly every OpenAI repository. It aims to highlight only high-confidence issues, preserving human attention for more complex tasks. Structured planning is essential for managing complicated projects, as demonstrated by the rapid development of the Sora Android app in just 28 days. The episode emphasizes using Git worktrees for parallel development, allowing multiple coding tasks without conflict. The new GPT-5.2 model has shown significant improvements, solving previously challenging bugs with reduced thinking time. Human oversight is becoming the real limiting factor in AI productivity, shifting the focus from code generation to decision-making about what should be built. Contextual prompting is critical; Embiricos advises providing specific backgrounds when interacting with Codex. He also highlights an interesting point: treating AI with respect can have positive implications for human interactions. The future of AI tools leans toward integration within existing workflows, allowing for seamless interactions without the need to switch contexts. This evolution in AI tools requires an adaptable interface to maximize their capabilities.