When AI Writes the Code, CI and Code Review Become the Bottleneck

For years the assumption in software teams was that writing code was the slow part. AI coding tools are testing that assumption. When code becomes cheap to produce, the constraints move to the places that were always there but rarely limiting: testing, review and deployment.

A real example from a company that measured it

Anthropic, the maker of Claude, recently described this shift in its own engineering organisation. In a post about its continuous integration (CI) system, it reported that CI workload grew about 25-fold over six months. Its engineers were shipping roughly eight times as much code per quarter as in earlier years, with Claude writing around 80% of it, and the number of tests grew about tenfold with only a small change in headcount.

These are the company's own figures about its own setup, so they should not be read as a universal benchmark. But the pattern they describe matches what many teams report: more, smaller changes, arriving around the clock, each triggering a full pipeline.

Why AI-assisted teams strain their pipelines

AI agents behave differently from human developers:

• More pull requests, each smaller. Agents commit in fine-grained steps.
• No working hours. Work arrives continuously, not in a daytime wave.
• More tests. Generating tests is cheap, so test suites grow quickly.
• Repeated runs. An agent that iterates on a failing check triggers the pipeline again and again.

A CI system sized for a team of humans can crumble under this. Queues grow, builds slow down, and developers wait longer for feedback than before they adopted the tools.

The review bottleneck

The second squeeze is human attention. If a person is asked to approve three times as many changes, reviews tend to become skims. That is a real risk, because reviewers are the last check before code reaches production. Anthropic itself said code review had become a bottleneck when it introduced an AI-assisted review feature in March, which tells you the problem is being felt even by those building the tools.

What to do in a normal-sized team

You do not need a huge engineering organisation to prepare. These changes help at any scale.

1. Run only the tests that matter. Test impact analysis, which maps a change to the tests it could affect, avoids rerunning the whole suite on every commit. Most modern CI tools support some form of change-based filtering.
2. Cache aggressively. Dependency, build and test caching removes a lot of repeated work.
3. Set budgets for agents. Limit how many pipeline runs an automated agent can trigger per task, so a stuck loop does not burn through your CI minutes.
4. Keep changes small and described. A short pull request with a clear description is faster to review, whether a human or a tool wrote it.
5. Decide what needs a human. Authentication, payments, data migrations and permissions deserve careful human review. Low-risk changes such as copy edits can go through lighter checks.
6. Watch the metrics that matter. Track time from commit to green build, review wait time and the rate of changes that get reverted. Lines of code produced is the wrong number to celebrate.

Bottom line

AI coding tools do not remove the need for engineering discipline; they raise the price of not having it. A team that speeds up code production without upgrading its tests, pipeline and review habits will simply move the queue somewhere else.

If your team has started using AI assistants and builds feel slower than they used to, look at CI and review first. That is where the extra output has to go.