2026-06-12 11:54 UTC
I increasingly think we require full e2e black-box test coverage that can survive a codebase rewrite for serious (long-term usage) software. I had a coding agent do overnight what would have taken me weeks... how am I now supposed to trust it? The more it looks like it worked the scarier it is.
Replies (2)
-
@lungching@mastodon.social 2026-06-12 12:02
@awwaiid@mastodon.social Did you check that it isn't taking fractions of pennies and putting them in an unknown account? Or code that will start the construction of muscley Austrian androids?
-
@mistersql@mastodon.social 2026-06-12 12:29
@awwaiid@mastodon.social - Manually spot check - Run ordinary code linters on the tests or other "tests for your tests" tools - Ask a bot from a different model (claude reviewing claude tends to not find problems, but chatgpt reviewing claude or vica versa is more likely to find problems if there are any) - Instead of 100% coverage, get the code covered two or three times - by default bots will write tests that assert current behavior. You have to ask them to specifically write failing tests to find bugs.