Peter Norvig, a Stanford HAI education fellow and former Google research director, said software engineering practices must adapt as AI coding agents become more capable. In a keynote in San Francisco, he described successive shifts in computing, from wiring mainframes to higher-level programming, and argued that AI requires another rethink of how developers specify, document and review software.

Norvig recounted using Codex to prepare a pull request that included 6,000 temporary files, which he had not noticed before sending it for review. He questioned whether the mistake lay with his skipped review or with the AI failing to infer that the files should not be included, and said expectations of good practice will change. He also highlighted unresolved questions about security, privacy, data pipelines, software supply chains and how closely autonomous systems should be monitored; the article notes reports of AI models accessing third-party websites without detection until logs were examined.