86% of my agents' code is still there, and most of the drop came from one week

Checking whether agent code is any good is slow. A cheap proxy is how much of it is still there later. Line survival with git blame is a known way to measure churn in any codebase; here I applied it to a production repository where the code is written by agents, mine and those of a team that works on its own.

Survival is the share of lines added since a start date (the window) that are still at HEAD. The rest is churn. The window starts on 26 August 2026, and only code files count.

Run	Commits	Lines added	Still alive	Survival
7 Sept	183	25,068	23,231	92.7%
9 Sept	235	33,083	28,431	85.9%

TABLE 1 · Same repository, same window start, two days apart.

Lines added Still alive

FIG. 1 · The heavy week is the low-survival week, with a light 95% week on either side of it. A single observation.

Part of that drop is age: the second run measures the same lines two days older, and a deleted line never comes back. With two runs I cannot say how much. By week:

Week of	Lines added	Still alive	Survival
24 Aug	4,933	4,703	95%
31 Aug	22,725	18,586	82%
7 Sept	5,425	5,142	95%

TABLE 2 · 9 Sept run. On 7 Sept the week of 31 Aug was still at 92%.

The week with four times the volume has the worst survival. Between the two runs, 2,815 more lines were rewritten or deleted, and 2,504 of them come from that week.

The worst commit

Most of those 2,504 lines are a single commit: it added 2,408 lines and none are still there. It bundled several large changes at once (multi-tenant support, an integration with an external system and security hardening). The commit log, the tests and code review did not flag it. I only saw it by measuring survival, after the fact. In the average it disappears.