Anthropic's Testing Load Grew 25 Times in Six Months Because Claude Writes the Code

Anthropic published a rare set of internal numbers this week about what happened to its own engineering department once coding agents took over most of the typing. Over six months, the volume of automated checking jobs grew 25 times. Engineers shipped roughly eight times as much code per quarter as they did in the years from 2021 to 2025, with Claude writing about 80 percent of it, and the number of tests in the codebase grew tenfold. The engineering headcount barely moved.

The bottleneck this created is worth understanding even if you never write software. Continuous integration, usually shortened to CI, is the automatic checking that runs every time someone changes the code: it rebuilds the project and runs the tests to see whether the change broke anything. When a codebase gets big, running every test on every change becomes too slow, so teams use test impact analysis, a service that works out which tests a given change could actually affect and runs only those. That service is what buckled. Anthropic’s engineers patched it three times and reported that the fixes held for 70 days, then 29 days, then less than a day, before they gave up patching and redesigned it. “Writing code is no longer the constraint, and once PR review gets accelerated, CI starts feeling the pressure,” engineer Sachin Malhotra wrote. His advice to other teams was to assume their systems will face 25 times their current load within two quarters, and to build for ten to twenty times what seems necessary.

What is behind this

This is the clearest published example so far of a pattern worth naming: automating a step does not remove the bottleneck, it moves it downstream. Code was the slow part, so it got automated, and now the slow parts are reviewing the code and proving it works. Notice that the tests grew tenfold too, which means the verification work did not shrink when the writing sped up. It grew with it. A fair caveat: these are self reported figures from a company that sells the coding agent in question, published on its own blog, and “Claude wrote 80 percent of the code” is a measurement whose definition Anthropic controls. The operational details, three failed patches and a rebuild, read as genuine engineering rather than marketing, which is what makes the post interesting.

What this means for you: If you work anywhere that is introducing AI tools, this is the question to bring to the meeting: once this step gets faster, what is the next step, and can it take the volume? The answer is usually a review process, an approval queue or one overloaded colleague, and those do not scale by buying more computers. If you write software, the practical version is to look at your checking pipeline before your team leans harder on agents, since a pipeline that is comfortable today can fall over at five times the volume. And if you are just curious about how much AI actually writes production code at a frontier lab, this is a real number from the inside, offered with the appropriate pinch of salt.

Sources

• Anthropic: Agentic coding is straining CI, here’s how we scaled test impact analysis
• Analytics India Magazine: Anthropic’s CI jobs grew 25x in 6 months as Claude wrote 80% of code