Coding agents add mocks in 36% of test commits where humans add them in 26%

The gap is 10 percentage points, and the post's author is careful about what that licenses. The author wrote: "Ten percentage points on a measured dataset is a clear signal; it is not evidence that agents mock everything in sight." ^[6] As a ratio, agent test commits add mocks about 1.4 times as often as human ones ^[7].

The sample limits how far that ratio travels. Agents wrote about 4% of the 1.2 million commits ^[8]. The post does not say how many of those are test commits or which languages and agents the set covers. For the 36% to hold in your repo, your agents, your language and your mocking history would have to look like the public sample ^[3].

The causal story is the author's. The paper, as the post describes it, measured rates of test-double use and brittleness ^[1][9]. The remedy is the author's inference: the fix lives at the convention layer, because review cannot keep pace ^[12]. The mechanism offered is that the codebase is the specification. "Agents mock what the codebase mocks," the author wrote ^[10].

The case against review rests on an illustration. A reviewer who catches eight mocks in a diff of twenty tests has shipped twelve, and those twelve become examples the next session samples ^[11]. The numbers are tidy, as hypotheticals tend to be.

In our view, the remedy is cheap enough to try before the evidence is complete. The first step is to measure: the share of test commits that add a test double, split by agent and human authors. That is the comparison the paper made ^[3][4]. If your own gap comes out near zero, a policy can wait. If your suite is already mock-heavy and agents write tests at volume, we think the post's case holds, and a written convention gives the reviewer something to enforce.