# Anthropic replaces git internally after its merges hit 30,000 a day, from 2,500 in May; a team of seven built Scion, the in-house system > Source: > Published: 2026-10-14 15:30:11+00:00 # Anthropic replaces git internally after its merges hit 30,000 a day, from 2,500 in May; a team of seven built Scion, the in-house system Anthropic outgrew git at 30,000 merges a day, so a team of seven built Scion, an in-house system it says isn't for general use. Three more stories inside. Anthropic's engineers explain why git stopped working for a repo where nearly every client is an agent, and what they built instead. It leads today because it's the first detailed account of agent-scale coding breaking the tools around it, and because, by Anthropic's own account, the fix isn't for anyone else yet. Also today: Claude Code 2.1.299 fixes a sandbox escape and adds idle compaction, GitHub Copilot adds Google's Gemini 4 Argon across its IDEs, and a one-person leaderboard called Parsimony ranks coding models by how much code they add per solved task, a day after a Harvard study found agents write 30% more code without shipping more software. **In this issue** • [Anthropic replaces git internally after its merges hit 30,000 a day, from 2,500 in May](https://www.thenewway.ai/anthropic-replaces-git-internally/#anthropic-replaces-git-internally-after-its-merges-hit-30000-a-day-from-2500-in-may) • [Claude Code 2.1.299 fixes a sandbox escape and adds idle compaction](https://www.thenewway.ai/anthropic-replaces-git-internally/#claude-code-21299-fixes-a-sandbox-escape-and-adds-idle-compaction) • [GitHub Copilot adds Google's Gemini 4 Argon across its IDEs](https://www.thenewway.ai/anthropic-replaces-git-internally/#github-copilot-adds-googles-gemini-4-argon-across-its-ides) • [Parsimony ranks AI coding models by how much code they add per solved task](https://www.thenewway.ai/anthropic-replaces-git-internally/#parsimony-ranks-ai-coding-models-by-how-much-code-they-add-per-solved-task) ## What changed this week ### Anthropic replaces git internally after its merges hit 30,000 a day, from 2,500 in May Anthropic uses Claude to write software "far more than almost any other company," its post says, and the scale broke its git setup. In May the company averaged 2,500 code changes merged per day; by September it merged over 30,000 per day. Nearly everything talking to its source control is an agent or an automated job, not a person, Anthropic says. It ran several hundred internal git mirrors to cope, with a main-branch history near 100GB. Its fix: [Scion](https://www.anthropic.com/engineering/redesigning-source-control?ref=thenewway.ai), an in-house version-control system, plus a code-review tool called Antique, built by a team of seven. The new setup draws 200x less CPU and memory than the old mirrors, Anthropic says. "All of the engineers who worked on this system would now rate Claude as better at writing code than them," the post adds. It isn't suited to general use, Anthropic says. If your repo isn't merging 30,000 times a day, none of this is for you yet. ### Claude Code 2.1.299 fixes a sandbox escape and adds idle compaction Claude Code's 2.1.299 release fixes a sandbox escape: a shell command could write into a plugin folder, whose code Claude Code reloads and runs outside the sandbox. It also stops Claude Code from sending a third-party gateway's API key to Anthropic on feature-flag and startup requests, and closes a macOS symlink bypass of denied paths. New: [idle compaction](https://github.com/anthropics/claude-code/releases/tag/v2.1.299?ref=thenewway.ai) trims a 200K-plus-token conversation left idle around 54 minutes, before its prompt cache expires, so your next message doesn't reprocess it. ### GitHub Copilot adds Google's Gemini 4 Argon across its IDEs GitHub says Gemini 4 Argon, Google's newest model, is now generally available in Copilot, across VS Code, Visual Studio, the CLI, JetBrains IDEs, Xcode, and Eclipse. Copilot Pro+, Max, Business, and Enterprise plans get it, billed by usage at Google's list price. [GitHub's changelog](https://github.blog/changelog/2026-10-13-gemini-4-argon-in-github-copilot?ref=thenewway.ai) says early testing solved more tasks than earlier Gemini models on every coding benchmark, with its biggest gains in the CLI. No benchmark names or scores are given. Check your model picker. ### Parsimony ranks AI coding models by how much code they add per solved task A new leaderboard called Parsimony, a one-developer open-source project by someone posting as lasu, scores AI coding models by how much code they add per solved task, not just whether the fix passes. It takes the patches [SWE-bench Verified](https://parsimony.lasu.dev/?ref=thenewway.ai)'s maintainers already publish, all produced by the same simple agent, and counts code units added (statements, calls, names, literals), net of deletions. On leaderboards with 26-plus models, more code tracks weakly with more solved tasks (rank correlation 0.33–0.60). ## Also worth your time Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one. *The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.*