Anthropic just replaced Git with their own system called Scion after $CLAUDE started merging 30,000+ code changes per day (up from 2,500 in May).
Why it matters:
- Their main repo hit 100GB. Testing infrastructure was doing 25M checkouts/day. Git couldn't handle it.
- Scion is 200x more efficient on CPU/memory. Checkouts went from minutes to seconds. Freed up 25% of CI capacity.
- Claude Mythos Preview literally rewrote parts of Scion to use 20x less memory and run 4x faster. The AI built the infra to handle its own output.
The meta shift: Coding isn't the bottleneck anymore. Version control, builds, and testing infra designed for humans can't keep up with AI agent velocity.
Anthropic isn't open-sourcing Scion but wants to work with GitHub/GitLab on solutions. They expect more teams to hit these limits in 6-12 months.
This is what scaling AI dev looks like when the agents outpace the tooling.
Why it matters:
- Their main repo hit 100GB. Testing infrastructure was doing 25M checkouts/day. Git couldn't handle it.
- Scion is 200x more efficient on CPU/memory. Checkouts went from minutes to seconds. Freed up 25% of CI capacity.
- Claude Mythos Preview literally rewrote parts of Scion to use 20x less memory and run 4x faster. The AI built the infra to handle its own output.
The meta shift: Coding isn't the bottleneck anymore. Version control, builds, and testing infra designed for humans can't keep up with AI agent velocity.
Anthropic isn't open-sourcing Scion but wants to work with GitHub/GitLab on solutions. They expect more teams to hit these limits in 6-12 months.
This is what scaling AI dev looks like when the agents outpace the tooling.