OpenAI Leader: Managers as 'Surgery Support'
OpenAI's Head of Platform Engineering, Sherwin Wu, argues that in an AI-native world, engineers will act as "wizards" managing agent fleets. The role of managers will shift to being "surgery support," focusing on unblocking the top 10% of issues that automated agents can't handle.
The shift from engineer-as-coder to engineer-as-wizard is already happening inside OpenAI. Roughly 95% of its engineers use Codex daily, with some managing fleets of 10 to 20 parallel AI agents. This has resulted in a 70% wider gap in pull requests between those who heavily use AI tools and those who don't. This new paradigm redefines productivity metrics. At OpenAI, code review times have been slashed from an average of 10-15 minutes per pull request to just 2-3 minutes, as every PR is now reviewed by Codex before a human sees it. One team is even maintaining a codebase written entirely by AI, with zero human-authored code. The core advice for developers building AI products is to anticipate the technology's trajectory. Wu warns against building complex "scaffolding"—like vector databases or agent frameworks—to compensate for current model weaknesses, as future models will likely absorb these functions. The focus should be on designing for where the models are headed in the next 12 to 24 months. This operational model aligns with Apple's long-standing strategy of vertical integration and on-device processing. By embedding AI capabilities directly into hardware like the Apple Neural Engine, the company minimizes latency and enhances user privacy—key advantages for running sophisticated, multi-agent systems efficiently and securely. For leadership, the challenge shifts from direct task management to orchestrating human-AI collaboration and fostering a culture of continuous learning. As AI agents handle more of the software development lifecycle, from planning to testing, managers must focus on strategic alignment, ensuring AI efforts deliver business value and that engineers develop the skills to manage these intelligent systems effectively.
