AI copies whatever it sees — so every deviation from the team’s standard breeds more of itself.
The team friction A codebase with mixed conventions used to cost a little friction. Now it costs compounding drift: an agent picks an existing file as its example and faithfully reproduces whatever’s there — including the team’s worst habits. One inconsistent corner becomes the template the next agent imitates, and the next. The codebase’s average quality becomes its destiny.
The pattern The team holds a single, high, consistent standard and keeps the codebase clean enough that the nearest example is always a good one. Shared conventions, linters, and a genuinely excellent reference area aren’t polish — they’re how the team steers every agent at once without reviewing every line. You raise the floor so the agents raise themselves to it.
Why it works with AI An agent’s strongest signal is the surrounding code, not the style guide. So consistency stops being aesthetic and becomes control: the team programs agent behaviour by curating what agents imitate. Excellence in, excellence out.
Watch-outs This needs collective discipline — one person’s “I’ll just match the old bad pattern” reintroduces the rot for everyone. And a written standard the codebase contradicts loses; agents follow the code.
Related Cleanup time after agents · ADRs — record the why · Agent-written shared docs