LectureRechercheRadarCadre d'investissement
Connexion / Inscription
← Retour à la lecture
ActualitésSujet · AnthropicMatériel1 min de lectureRédaction DisconfirmAI

AI Engineer Podcast: Anthropic Scales Code Merges to 30,000 Daily via Claude Review

The AI Engineer podcast releases an episode discussing how Anthropic scaled its software development throughput. Guests Manish Kapur (Sonar) and Addy Osmani (Anthropic) cover the shift from human-paced reviews to AI-assisted workflows, specifically detailing how Anthropic moved from approximately 2,500 to 30,000 merged code changes per day using Claude for first-pass review backed by verification evidence.

2026-10-29 07:00 GMT+8

The AI Engineer podcast has published a new episode titled "How Anthropic Merges 30,000 Code Changes a Day," featuring Manish Kapur from Sonar and Addy Osmani from Anthropic. The discussion centers on the challenges posed by coding agents generating pull requests faster than traditional human review processes can handle.

According to the episode description, Anthropic scaled its internal software factory from about 2,500 to 30,000 merged changes per day. This increase was achieved by implementing Claude for first-pass reviews, supported by real verification evidence. The guests also discuss the concepts of "light" versus "dark" software factories and findings from Sonar's State of Code survey regarding where next efficiency gains might come from.

The episode covers technical controls such as permissions and sandboxes, and concludes with a lightning round on labeling agent-written PRs and determining when it is safe to merge. This resource offers a specific look at how a major AI organization is restructuring its engineering validation loops to accommodate high-throughput AI-generated code.

Sources

Connexion / Inscription
LectureRechercheRadarCadre d'investissement
youtube.com
↗
Sujet

Anthropic

【背景】Anthropic披露其内部工程实践:通过引入Claude进行自动化代码审查并重构底层源码管理系统,该公司将每日代码合并量从2026年5月的约2,500次激增至9月的超过30,000次。这一变化标志着软件开发瓶颈已从“编写代码”转移…

4 étapes1 piècesMis à jour 2026-10-30 13:03

Comment on en est arrivé là

  1. 2026-05-01Anthropic内部每日代码合并量约为2,500次,作为基准线。finance.biggo.com ↗
  2. 2026-09-01每日代码合并量突破30,000次,增长逾12倍;同时Claude Code团队报告PR回滚率下降超40%。x.com ↗
  3. 2026-10-13Anthropic发布工程博客,宣布内部Monorepo从Git迁移至自研系统Scion,以解决高并发下的性能瓶颈。note.com ↗

Ce qui reste non résolu

  • Scion系统的非开源策略是否会阻碍其在更广泛开发者社区中的采用,或导致Anthropic内部工具链与行业标准脱节?
  • 随着AI生成代码量的指数级增长,现有的自动化验证框架能否持续保证安全性,特别是在处理复杂逻辑漏洞时?
  • 高昂的多智能体审查成本(如每次21美元)是否意味着只有大型科技公司才能负担得起这种规模的AI辅助开发,从而加剧行业分化?
Lire le sujet complet →
Continuer la recherche →