Detailed Analysis
A report from Chinese tech outlet 36Kr describes an AI-driven coding project in which Anthropic's Claude model reportedly rewrote millions of lines of code over an 11-day span, triggering significant backlash online. While the underlying article is only available as a truncated headline via news aggregation, the framing itself is notable: it points to a large-scale, autonomous or semi-autonomous refactoring effort executed by Claude at a pace and volume that would be effectively impossible for human engineering teams working alone, and it suggests that the scale and speed of the change—rather than any single technical failure—is what generated public controversy.
This kind of story fits squarely within a broader pattern that has emerged around Claude Code and similar agentic coding tools throughout 2025 and into 2026. Anthropic has increasingly positioned Claude not just as a code-completion assistant but as an agent capable of planning, executing, and iterating on substantial engineering tasks with minimal human oversight—reviewing entire codebases, proposing architectural changes, and applying them across many files simultaneously. Millions of lines of code being touched in under two weeks is consistent with the kind of throughput these agentic systems are now marketed as capable of achieving, especially when deployed against legacy systems, large monorepos, or migration projects where repetitive, pattern-based transformations dominate the work.
The "public outrage" angle is the more consequential part of the story, and it echoes concerns that have surfaced repeatedly as AI coding agents move from writing snippets to executing sweeping, autonomous rewrites. Developers and engineering organizations have voiced anxiety over several intertwined issues: the difficulty of meaningfully reviewing AI-generated changes at that scale, the risk of subtle bugs or security vulnerabilities being introduced silently across a massive surface area, questions about who bears responsibility when an autonomous rewrite breaks production systems, and the labor implications of a single AI system performing work that might otherwise require dozens of engineers over months. There is also a cultural dimension specific to how such stories circulate in Chinese tech media, where AI capability demonstrations are often received with a mixture of awe and skepticism about hype, corporate overreach, and the erosion of engineering craftsmanship.
More broadly, this episode reflects the accelerating tension between AI capability gains and institutional readiness to govern them. As Claude and competing models from OpenAI, Google, and others push further into autonomous software engineering—sometimes described as "agentic coding" or "vibe coding"—the industry is grappling with whether existing code review practices, testing infrastructure, and organizational trust models can keep pace with AI systems that can generate and modify code orders of magnitude faster than humans can verify it. Incidents like this one, regardless of the specific technical details, function as flashpoints in an ongoing public debate about how much autonomy to grant AI coding agents, how to maintain accountability and auditability in AI-assisted software development, and what guardrails—technical, procedural, or regulatory—should accompany the deployment of increasingly capable coding models like Claude in production environments.
Read original article →