Skip to main content

Claude's Code Overload Nearly Broke Anthropic's CI System

When AI Writes the Code, the Pipeline Pays the Price

At Anthropic, something unusual is happening. 80% of the company's code is now written by Claude, their own AI model. Engineers are delivering eight times more code per quarter than the average from 2021 to 2025. Even code reviews and merge approvals are largely handled by the AI.

But here's the catch: AI doesn't code like a human. It submits tiny pull requests all day long, flooding the system with test cases—ten times the usual number. The result? CI tasks skyrocketed 25-fold in just six months, nearly crashing the entire continuous integration (CI) system.

Three Patches, Three Failures

Anthropic tried to patch the problem. They added more cores, sharded the workload, and even forced daily restarts. None of it worked. The system kept choking under the relentless pace of AI-generated code.

Finally, they turned to Claude for advice. The AI suggested a complete rebuild: a distributed, stateless architecture. That did the trick. The CI overload was solved—not by tweaking the old system, but by rethinking it entirely.

The Real Bottleneck Isn't Code—It's the Pipeline

This incident sends a clear signal. The question is no longer whether AI can write code. It's whether our software engineering systems can keep up. When code generation is no longer bound by human rhythms, the first things to break are the pipelines designed for human habits.

Think about it: if AI can produce code faster than we can test and merge it, what's the point of faster coding? The bottleneck shifts from writing to integrating. And that's a problem we haven't fully solved yet.

Key Points:

  • Claude writes 80% of Anthropic's code, boosting output eightfold.
  • AI's fine-grained PRs caused CI tasks to surge 25 times in six months.
  • Three patches failed; a distributed stateless rebuild finally fixed it.
  • The bottleneck in AI-driven development is now the CI/CD pipeline, not coding speed.