Budget the team’s time to tidy what agents leave behind — or the mess compounds into everyone’s.
The team friction Agents produce working code fast, and “working” hides a lot: dead branches, near-duplicate helpers, inconsistent naming, tests that assert nothing. Each person’s leftovers look minor alone. But they land in the shared codebase, where they compound — the next agent copies the mess as precedent, and soon everyone is slower in a codebase no single person chose to degrade.
The pattern The team explicitly budgets cleanup as part of the work, not an optional afterthought: a slice isn’t done until its agent output is tidied to the team’s bar. Make it a standing, shared expectation — refactor passes, dead-code sweeps, a “leave it cleaner” norm everyone is held to.
Why it works with AI Generation is cheap, so the team produces more raw material than ever — and absorbing it (reading, pruning, normalising) is the expensive half that doesn’t get cheaper. Naming the cleanup keeps the expensive half from silently becoming everyone’s tax.
Watch-outs Don’t let “cleanup” balloon into endless gold-plating. Tie it to the team’s standard, then stop.
Related Consistent, excellent codebase · Shorter iterations · Trap: measuring AI adoption by tokens spent