You own the failure model — agents don't grade their own homework

Now the part that matters most in the agent era, and the part most easily skipped.

A test only catches what it was written to check. A suite is a lower bound on correctness — proof that the cases you thought of work, and silence about everything else. So the real skill isn’t writing tests; it’s modelling how the thing can fail and turning that model into checks. What must never happen? A workspace must never exceed its seat limit. A teammate must never be invited twice. Those are your assertions about the world, and they’re the most valuable lines in the codebase.

Which is exactly why you cannot let the agent write them alone. If the same agent writes both the implementation and the tests, it tests the cases it happened to consider — and its blind spots land identically in both. It will hand you green checks and a bug, because it graded its own homework. The failure model — the “what must never happen” — has to be owned by a human, or at minimum specified by one and verified independently. This is the single habit that separates disciplined agent orchestration from hopeful prompting.