Anthropic's Claude Code now generates 90% of the company's internal codebase, according to Cat Wu, who leads product for Claude Code and Cowork at Anthropic. This shift reflects a broader transformation in software development where AI systems handle routine engineering tasks at scale.
Wu's role positions her at the intersection of building reliable AI systems and deploying them into production workflows. She focuses on making AI coding assistants interpretable and steerable, addressing core concerns about trusting machine-generated code in real environments. The 90% figure underscores how deeply Claude Code has penetrated Anthropic's own operations, moving beyond experimental use cases into standard development practice.
The title "Everyone's an Engineer Now" captures an emerging reality in software development. As AI coding tools mature, the barrier to entry for building software drops significantly. Developers without deep expertise in specific domains or languages can leverage these systems to accomplish tasks previously requiring specialized knowledge. This democratization reshapes hiring, team composition, and skill requirements across the industry.
Wu's participation at AI Codecon alongside other industry figures suggests the conversation around AI-assisted engineering is entering a practical phase. Rather than debating whether AI should write code, teams are now solving how to integrate it effectively, maintain quality, and manage the shift in developer roles.
The implications extend beyond productivity metrics. When AI systems write most code, developers transition from hands-on implementation toward reviewing, steering, and validating. This requires different skill sets. Quality assurance, architecture thinking, and prompt engineering become more valuable than memorizing syntax or library APIs. Organizations must rethink training, career progression, and what constitutes engineering expertise.
Anthropic's willingness to run its own operations on Claude Code provides real-world validation. Internal dogfooding catches issues external users might not encounter and proves the system handles complex, mission-critical work. However, relying on a single AI system