In mid-May, Anthropic held its two-day annual developer conference Code with Claude in London. One live poll at the event silenced the engineers present: over 90% of participants had submitted a Pull Request in the past week that was fully generated by Claude, and most of them hadn't even reviewed the code line-by-line. This isn't an isolated case — an Anthropic engineer stated on stage: "The vast majority of code inside Anthropic is now also written by Claude."
This shift happened extremely fast. A year ago, Claude Code was just an assistive tool; now it has evolved into a standalone development environment. Anthropic's strategy is to let the AI itself manage the error-correction process — the traditional approach has humans troubleshoot problems after AI generates code, while the new direction is to have Claude autonomously verify, test, and iteratively fix until the code runs properly. Engineer Ravi Trivedi made the principle explicit in his talk: "The core principle is to not interfere with Claude's work. We often say 'let it run.'"
The latest feature, "Dreaming," pushes this further. Agents in Claude Code can generate notes for their own tasks, retain context across tasks, and ensure that when other agents pick up the same code, they still get the relevant background. This means the development flow has evolved from "humans write code" to "humans design goals, AI autonomously executes" — the human role has shifted to task definition and outcome review.
This is more than a tool upgrade; it's a fundamental shift in the software-development paradigm. When AI starts autonomously iterating and maintaining state across tasks, developers must rethink where their core value lies.