Anthropic is testing Claude Code for daily maintenance of its own applications. According to The Decoder, over a few weeks the tool generated 388 change-merge requests, and 46% of them were merged after human review.
Tasks included fault finding using fuzzing and removal of unused code. The inventor of Claude Code, Boris Cherni, described the result as 'early signs that this might be possible.'
The significance of the experiment lies in testing a scenario where AI performs not a one-off code generation but ongoing technical work within the company. However, the available information is presented as metadata from a single independent source, leaving questions about the quality of changes and the distribution of tasks.
editorial commentary
Why it matters
If such metrics persist across a broader set of tasks, the next observable signal will be the publication of data on quality, rejections, and validation costs. The main uncertainty is the absence in the available materials of a detailed experimental methodology and independent confirmation.