90% of Claude's Code Written by AI

This represents a fundamental change in how AI labs operate internally. Instead of hiring armies of engineers to write model training code, Anthropic is using Claude to write Claude. The implications for training data are profound—the feedback loop becomes recursive, where AI-generated code creates the systems that train the next generation of AI. But this creates new evaluation challenges. When AI writes AI, failure modes compound. If Claude has coding blind spots, those get embedded into the next version of Claude. Labs know this, which is why they still need external human evaluation—not to write code samples, but to catch what automated systems miss. The data labeling opportunity shifts from 'write good code' to 'evaluate AI-generated code for subtle problems.' This requires annotators who can identify when code works but is inefficient, passes tests but has edge-case failures, or solves problems but introduces security vulnerabilities. These are higher-skill, higher-value annotation tasks. The timing aligns with broader industry momentum. Developer forums show Anthropic's Claude Agent SDK emerging as the leading framework for building agents. If this 90% figure is accurate, Anthropic is betting heavily that autonomous systems will handle most future development work—both internally and as products customers deploy.