When AI writes its own code: Who watches the watchers at Anthropic?
In the world of modern development, we are used to using AI to write code snippets, but at Anthropic, they have taken it one step further – much further. Today, over 80% of the code embedded in the company's codebase is written by Claude [2, 10]. If you are wondering who checks all this code, the answer is a fascinating combination of automation, artificial intelligence, and deliberate human involvement.
The quiet revolution of autonomous code writing
The data from Anthropic shows a dramatic shift in the pace of work. The company's engineers are now shipping 8 times more code every quarter compared to previous years [2, 10]. This does not mean they are working 8 times harder, but rather that they are using AI to perform the "grunt work." Claude doesn't just write the code; it is also an active participant in its optimization and improvement process [10].
But with such a volume of machine-generated code, the question of security and quality arises. How do you ensure the code doesn't contain critical bugs? This is where Anthropic's automated Code Review process comes into play [6, 9].
What does the new review process look like?
Anthropic has implemented a system where every code change undergoes scanning by dedicated AI agents [6]. The process works like this:
- •Automated scanning: Every Pull Request is scanned by AI agents looking for bugs, design issues, and security vulnerabilities [6].
- •Retrospective analysis: Internal analyses have shown that had this tool been used in the past, a third of the bugs that reached Production could have been prevented [10].
- •Human involvement: Despite the automation, humans remain in the loop. The higher the risk of the change, the greater the human involvement in the final review and approval [1, 3].
Data that indicates a trend shift
Data from recent months shows an interesting trend: while the number of changes entering the codebase doubled between June and September 2026, the rate of changes that required a Rollback dropped by more than 40%. This is an amazing figure, indicating that the combination of AI-written code and AI-reviewed code makes the process more stable and reliable than human work alone.
In addition, Anthropic recently published new metrics to track the pace of AI development, including the level of autonomy of AI agents and the scope of their oversight [3, 7]. As of today, about 30,000 agents are operating in the company's internal systems at any given moment, with the system flagging on average one or two out of every thousand for in-depth human review [1, 3].
Summary and conclusions
We are at a turning point where AI not only helps us write code but also manages its quality control. Anthropic's model proves that it is possible to significantly increase output without sacrificing code quality, provided there is a smart oversight system based on the same technology.
For developers and organizations, the message is clear: the future of development is not just in AI that writes code, but in AI that builds the testing and security pipeline around it. As these tools become more accessible (like the launch of Code Review for Enterprise customers), we will see more companies adopting this approach to stay competitive.
Further reading: