The era of human-centric version control is cracking under the weight of autonomous agents. In May 2026, Anthropic merged roughly 2,500 code changes daily into its monorepo. By September, that volume had surged 12x to 30,000 merges per day, driven almost entirely by Claude agents writing, testing, and shipping code. The infrastructure that served human engineers for two decades simply could not keep up with the machine-scale velocity.

Scion: A Radical Rebuild

Anthropic’s seven-person team spent five months building Scion, a custom version control system designed from scratch for agent-scale development, alongside a companion tool called Antique. They migrated their entire monorepo—nearly 100 GB of main-branch history—off Git and onto this new system. Scion inverts Git’s core assumption of offline-capable full copies; instead, it assumes clients are always online, keeping heavy computation server-side while clients maintain only bounded local caches. The performance gains are stark. Anthropic reports that Scion uses approximately 200x less CPU and memory than their previous fleet of several hundred Git mirrors. CI checkouts dropped from minutes to seconds, representing a two-order-of-magnitude speedup. Perhaps most telling, Claude itself rewrote the Git-to-Scion ingestion pipeline, using 20x less memory and running 4x faster than the handwritten prototype while discovering an algorithmic optimization for ancestry queries that human engineers had missed.

The Self-Referential Loop

This is a story about AI building its own infrastructure. Claude agents generated the code volume that broke Git, then helped build the replacement, and now monitor Scion’s health through "superfan" bots watching component telemetry. The Register bluntly framed the bottleneck as self-inflicted: Claude created the load, and Claude fixed the tooling. This shifts the ROI calculation for developer infrastructure, as agents can now design and implement the systems that support their own execution.

GitHub Rebuilds in Place

While Anthropic had the luxury of a clean-sheet design, GitHub is rebuilding its Git backend in place to preserve backward compatibility. Principal engineer Brian Celenza revealed that pushes grew 4.9x year-over-year to 3.35 billion per month, with 7.38 billion commits in September 2026 alone. GitHub’s new architecture decouples storage from compute, showing up to 35x higher write throughput in internal benchmarks, but unlike Scion, it must support the existing Git protocol and ecosystem.

Key Takeaways

  • Git’s distributed model, designed for human teams, fails at agent scale due to mirror sync issues and merge queue bottlenecks.
  • Anthropic’s Scion achieves 99%+ cache hit rates by treating clients as always-online, API-first consumers.
  • GitHub is adapting its backend to handle 14 billion projected commits for 2026, prioritizing compatibility over radical redesign.
  • For most teams, the bottleneck remains human review workflows, not Git itself, but infrastructure audits are now critical.

The Bottom Line

Git isn't dead for the 99.9% of teams, but the ceiling for human-speed collaboration is closing fast. If you are running coding agents, your infrastructure was built for humans, and it will break before your code does.