Decision record // Quality
When do tests start paying for themselves?
Everyone agrees tests are good and nobody agrees when to start. The usual advice was written before a machine was producing most of the code, which is why it now points in the wrong direction.
The call
Not on day one, and considerably earlier than most founders think. The trigger is not a coverage number. It is the first time you are afraid to change something, or the first time an agent writes code you are not going to read line by line.
When each one is right
Not yet
- The interface is still moving weekly, and every test you write is a test you will delete.
- You would be testing the framework rather than any decision you made.
- You are chasing a coverage percentage instead of answering a specific fear.
Now
- Money or data loss is on the line: payments, webhooks, migrations, anything that deletes.
- You are enforcing a boundary. Can tenant B read tenant A's rows. Can a signed-out visitor reach this route.
- You have now fixed the same bug twice.
- You are about to change something and you notice you are nervous. That nervousness is the most reliable signal there is, and it beats every rule of thumb.
What changed, and why the old answer is wrong
If an agent is writing your code, tests are the mechanism that makes its output verifiable. Without them your only quality control is reading every diff by hand, forever, which does not scale and which nobody is honestly still doing by week three. With them you can hand the agent a failing test and let it work until the test passes, which is a fundamentally different and much better loop. Tests stopped being mainly about catching regressions and became the specification you hand to the thing writing the code. That is why the honest answer to “when” is now earlier than it was two years ago, and why teams generating a lot of code with no tests are accumulating a debt they cannot read their way out of.
Decisions like this one, on your product
This page is a general answer. Yours has constraints I don't know about. The Founding Engineer is 8 weeks with 5 founders where we make these calls on your actual codebase, and your code gets reviewed every week.
New decision records as I write them
One question, one answer, no hedging. Leave your email and I'll send the new ones. No schedule promises, unsubscribe anytime.