Claude now writes 80% of Anthropic's own code, and it broke their CI
Anthropic's engineering team says code volume is up 8x since 2021-2025 with Claude authoring most of it, which pushed CI jobs up 25x in six months and forced a rebuild of how they decide which tests need to run.
Anthropic published an engineering post on September 14 with numbers on how its codebase has changed under agentic coding. Engineers now ship roughly 8x as much code per quarter as they did from 2021 through 2025, with Claude authoring 80% of it and playing a large role in reviewing and approving pull requests. The test suite grew 10x to keep pace, which pushed CI job volume up 25x in six months. That was enough to threaten the test-impact-analysis service that decides which tests need to run against a given change.
The bottleneck moved
The post says it directly: "writing code is no longer the constraint." Agentic coding speeds up writing and review, and the extra volume then floods the infrastructure downstream. Anthropic re-architected its test-impact-analysis system with headroom well beyond current load, sizing it on the assumption that load could grow another 25x within two quarters.
What other engineering orgs should expect
Frontier labs rarely publish internal telemetry on AI-assisted coding. The bottleneck described here (write fast, review fast, then overload CI) is what any org rolling out Claude Code or a similar tool at scale should plan for, and it's usually missing from the budget. If you track AI coding-tool ROI by lines shipped or PRs merged, watch CI queue times and review capacity alongside them. A spike in output says nothing about whether the infrastructure underneath is keeping up.