Anthropic published something this week that every engineering team building with AI coding agents should read closely.
Internally, Claude now writes roughly 80% of Anthropic's production code, and their engineers ship about 8x more code per quarter than they did from 2021 to 2025. That part isn't surprising anymore. What's more interesting is what broke under that load.
The numbers that matter
CI job volume grew 25x in six months. Test volume grew 10x. And the bottleneck didn't move to the AI model itself, it moved to everything downstream of it: code review, test selection, and CI infrastructure.
Anthropic's own framing is useful here: AI-generated code doesn't strain your model, it strains your systems. Agentic coding tools favor fine-grained, frequent commits, operate continuously instead of in human working hours, and multiply your test surface area far faster than a human team would.
What they did about it
Rather than re-running an entire test suite on every small change, Anthropic built smarter test-impact analysis, essentially figuring out which tests actually need to run for a given diff instead of brute-forcing full coverage every time.
Their advice to other engineering teams adopting agentic coding seriously: assume you'll hit roughly 25x load within two quarters, and design your infrastructure for 10 to 20x scale now, not later.
The real planning question
For any team rolling out AI coding agents, the question worth asking isn't whether the AI writes good code anymore. Most of the frontier models are good enough at that. The real question is whether your CI pipeline, your test suite, and your review process can survive several times the throughput without silently degrading in quality or grinding to a halt.
A few practical takeaways if you're scaling agentic coding on your own team:
Invest in test-impact analysis before you need it, not after your CI queue backs up for hours. Treat code review capacity as a first-class scaling constraint, since a human (or even an AI reviewer) bottleneck doesn't disappear just because code generation got faster. And expect your test suite itself to grow much faster than your team headcount, because agents write more tests than humans do by default.
None of this is a reason to slow down on agentic coding. It's a reason to make sure the plumbing around it can keep up.
Top comments (0)