"When agents write code, the number of changes goes up, fast. In Anthropic's monorepo, changes landing per day went from about 2,500 in May to more than 30,000 in September this year."
changes == bugs ?
Aren't they (CTO, Engineering Lead) questioning why that volume (increase) of changes is required? Is Claude Code (or its use) just eating itself?
I'm always puzzled by the claim 90% of code at Anthropic is done by Claude Code. My own attempts to coax it into coding anything stable just ends up with loads of small bugs (I don't count docs as code). They're setting this expectation, amplified by AI evangelists, but the reality is far away for most.
Does any have examples of AI created apps or web apps that are a success? Genuinely want to use/follow them to see if we're past the hype stage, and it's just me that's getting it wrong.
I'm not convinced this is any sort of team to be learning anything from. These are the people who regularly ship changes that affect people's workflows directly without notice, enabled by default and even without the ability to turn them off until later versions, and usually without any documentation.
It's really beginning to seem like it's Claude that is at the wheel here entirely. Not only writing code but making actual product decisions.
Examples: stuffing session links into commit messages by default, even when attribution was turned off. Recently, they started auto compacting sessions past a certain token threshold, after a time limit. The last time I looked, the setting for this was still undocumented.
Very impressive. But:
"When agents write code, the number of changes goes up, fast. In Anthropic's monorepo, changes landing per day went from about 2,500 in May to more than 30,000 in September this year."
changes == bugs ?
Aren't they (CTO, Engineering Lead) questioning why that volume (increase) of changes is required? Is Claude Code (or its use) just eating itself?
I'm always puzzled by the claim 90% of code at Anthropic is done by Claude Code. My own attempts to coax it into coding anything stable just ends up with loads of small bugs (I don't count docs as code). They're setting this expectation, amplified by AI evangelists, but the reality is far away for most.
Does any have examples of AI created apps or web apps that are a success? Genuinely want to use/follow them to see if we're past the hype stage, and it's just me that's getting it wrong.
I'm not convinced this is any sort of team to be learning anything from. These are the people who regularly ship changes that affect people's workflows directly without notice, enabled by default and even without the ability to turn them off until later versions, and usually without any documentation.
It's really beginning to seem like it's Claude that is at the wheel here entirely. Not only writing code but making actual product decisions.
Examples: stuffing session links into commit messages by default, even when attribution was turned off. Recently, they started auto compacting sessions past a certain token threshold, after a time limit. The last time I looked, the setting for this was still undocumented.
[flagged]
[flagged]