Agentic coding is straining CI. Here’s how we scaled test impact analysis at Anthropic
Anthropic’s AI-accelerated development dramatically increased code output, PR volume, and CI load, forcing a test-impact service through failed stopgap fixes before a redesign to a stateless, horizontally scalable architecture that could handle exponential growth.
MAIN POINTS
- Engineers shipped 8x more code, with Claude authoring 80%, sharply increasing PR and CI demand.
- Test counts grew 10x, driving a 25x rise in CI jobs and stressing the listener-selector system.
- Three patches bought shrinking relief: bigger machines, sharding, and daily restarts.
- The final redesign used an in-memory journal and stateless workers to enable horizontal scaling.
TAKEAWAYS
- Plan architectures for exponential growth, not linear scaling.
- Avoid single-instance critical services when AI can rapidly amplify workload.
- Keep process state external so services are easier to scale and recover.
- Instrument systems well enough for agents to monitor, diagnose, and tune them autonomously.