Anthropic details how it rebuilt CI after Claude-driven code volume grew 25x in six months

Anthropic published an engineering report explaining how it redesigned its continuous-integration (CI) test infrastructure after AI-assisted development caused CI job volume to explode across the company.

What's new

Anthropic says its internal CI job volume grew sharply as Claude took over more of its own software development. "This all led to a 25x increase in CI jobs over a six month period," the company wrote, noting engineers now ship roughly 8x as much code per quarter as they did from 2021-2025, with Claude authoring about 80% of that code and playing a large role in reviewing and approving pull requests. That volume overwhelmed the test-impact-analysis service Anthropic uses to decide which tests need to run for a given change, forcing engineers through three rounds of stopgap patches (bigger machines, then sharding, then daily restarts) before they redesigned the system on a horizontally scalable, database-backed architecture.

Context

Test-impact analysis exists to keep CI fast: instead of re-running an entire test suite for every change, the service maps code changes to the specific tests they could affect. That approach works at typical human commit volumes, but Anthropic found it broke down as Claude began generating and submitting far more changes, far faster, than a human engineering team would. The company frames this as a preview of a broader industry problem — CI, code review, and release tooling built for human commit cadences will need to be rearchitected as coding agents scale.

Why it matters

This is a rare look at how a frontier AI lab's own infrastructure strains under its own product's output. Anthropic's explicit advice to other engineering teams — to assume architecture will hit a 25x load within two quarters — is a concrete data point for any organization now deploying agentic coding tools internally, and a signal that AI-accelerated development is starting to bottleneck at the tooling layer, not just at model capability.