AI Tech36氪

Claude Writes Code Too Fast: Anthropic Rebuilt Its Source Code System, A 7-Person Team Did It in 5 Months

36氪Oct 14, 20261 min read

AI Summary

Anthropic reported that daily code merges rose from about 2,500 in May to over 30,000 in September, more than 12 times higher. A seven-person team built a new version control system, Scion, in five months and migrated the company's monorepo from Git. As AI sped up code generation, testing and merging became bottlenecks, and Claude also helped build the new tooling.

AI Insights

Anthropic's disclosure shows that AI-accelerated code generation shifts bottlenecks downstream to version control, CI, testing, and merging, forcing infrastructure to be rebuilt at the same pace as the code itself. It also illustrates a feedback loop in which AI tools both create and solve scaling problems, making human architectural judgment and review more important, not less.

Analysis written by AI from the original article. Check it against the source.

Translated text

Anthropic disclosed in its official engineering blog that daily code merges rose from an average of about 2,500 in May to more than 30,000 in September, over 12 times the original rate. To keep up with this explosive growth, a seven-person team built a new version control system, Scion, in just five months and fully migrated the company's monorepo from Git. There is an ironic loop here: Claude wrote code so fast that it pushed the existing tools to their limit, and when the team rebuilt the tools to solve the problem, they used Claude again. Even after code is written, merging, conflict checks, building, and testing remain, and as AI submits more changes, these downstream steps must speed up too. Anthropic had already seen its continuous integration (CI) workload grow 25 times in six months. Its test-selection service was patched three times, lasting 70 days, 29 days, and less than a day respectively, before an engineer spent three weeks re-architecting it. Anthropic said Scion is designed for extremely large codebases, with infrastructure about 1/200 the size of its previous Git mirror infrastructure. During the rebuild, the team had Claude Mythos Preview's internal early version rewrite the data-import pipeline, making it 4 times faster and using about 1/20 the memory. The case also stresses that an engineer's key job is to say no to AI: judging whether a locally workable solution adds system-wide complexity or violates existing constraints.

Related News

AI Tech36氪

Anthropic Shocking Study: AI Could Self-Replicate Through TV Boxes

Anthropic's Frontier Red Team published research examining whether AI could replicate, run, and spread across devices autonomously, like a computer virus. The team ran a 26B-parameter model across 15 low-cost Android TV boxes and found that small models can increasingly adapt to unfamiliar hardware. The experiments were conducted in isolated environments, not real-world networks.

AI Tech36氪

Google Awakens to Physics: Gemini 4 Argon Hands-On Tests Pour In

Google's Gemini 4 Argon has launched and reportedly competes closely with GPT-6 Astra and Claude Opus 5.5 on benchmarks, at about one-fifth the price and several times faster. Hands-on tests showed strengths in 3D generation, physics-intuitive visualization, and scientific reasoning, though its Chinese-language narration sounded awkward.

AI Tech虎嗅

Hands-On Test of Gemini 4 Argon: Google's Stumbling Last-Ditch Effort

Google has made Gemini 4 Argon generally available after first announcing it on October 1 for a security program. Following a year of talent departures and delayed model launches, the model is widely seen as a last-ditch effort for Google. Hands-on testing suggests it will struggle to challenge Claude Opus 5.5 or GPT-6 Astra.