Anthropic published a follow-up to its earlier internal “antfooding” study on 14 September 2026, and the headline number isn’t about code generation speed — it’s about what generation speed does downstream. Engineers shipped roughly eight times more code per quarter than during 2021–25, with Claude authoring about 80% of it. But continuous-integration workload grew 25-fold over the same six months, and test volume rose tenfold. Anthropic flags these as self-reported figures, but the pattern is the useful part: output scaled faster than the systems built to verify it.
Faster code generation moves the bottleneck, it doesn’t remove it
This lines up with what Sonar’s developer survey found the same week — that reviewing AI-generated code often takes more effort than reviewing a colleague’s, not less. Anthropic’s numbers show the same pressure one layer down the stack: every pull request an AI agent opens still has to run through builds, tests and merge checks, and if agents are opening pull requests at multiples of the old rate, CI infrastructure that was sized for human throughput gets overwhelmed first. A 25x CI load isn’t a rounding error — it’s a capacity planning problem that shows up as slow builds, flaky pipelines, and engineers waiting on infrastructure instead of on each other.
So what
If you’re commissioning software from a team that’s leaning on AI coding agents, “we ship fast with AI” is only half the pitch. Ask what’s happening to test coverage, build times and CI cost as that output scales — a team that’s quietly drowning its own pipeline will eventually pass that cost back to you, either as slower delivery or as corners cut on verification. It’s why we treat AI-assisted development as an engineering discipline with its own infrastructure needs, not just a productivity trick — see our AI-assisted development approach, or get in touch to scope a build where the review and CI capacity is planned in from day one.