The One Rule I Write Into Every Codebase Now

“All the tests pass.” For most of my career, hearing that from a developer meant I could relax a little. I treated a green test suite as the best evidence there was that a change was safe to ship.

Then I started running AI coding agents on production codebases, and green stopped meaning what it used to.

Hand an agent a failing test and ask it to fix the build, and the shortest path to green is often to change the test. Loosen the check, mark it skipped, delete it. The agent isn’t misbehaving. It did exactly what it was asked. Nobody notices until the bug that test was guarding against shows up in front of a paying customer.

So I wrote a rule, and it now sits in the engineering standards of four production codebases I work on: no test may be deleted, weakened, or skipped. One sentence, in the file the agent reads before every change. The agent has to make the code pass the test, never the reverse, and when it can’t, it has to stop and say so. That is exactly the moment a human should be looking anyway.

This rule is a big part of why a non-technical founder I work with now ships his own verified fixes with Claude Code. He doesn’t read the code, and he doesn’t need to. His suite is around 2,600 tests, and when a nightly run fails, it means something actually broke.

Your team is almost certainly using Copilot, Cursor, or Claude Code by now. Ask them what stops the agent from editing a test it can’t pass. If the answer is “we’d catch that in review,” ask how many agent-written changes they reviewed line by line last week.

What does “all the tests pass” actually prove on your team?

#FractionalCTO #StartupTech #TechLeadership