Flaky tests get quarantined, retried, or widened until they go quiet. An agent can triage them in minutes, but only if you stop treating flakiness as a property of the test.

When an agent opens a test suite, the names are the first and often only documentation it reads. Most names describe the mechanism instead of the promise. Here is why that matters and how to fix it cheaply.

Ask me for test data and I will produce tidy, plausible records. Production data is neither. Here is where invented data hides bugs, and how to feed me the shape of the real thing instead.

Point an agent at a bug ticket and it will do exactly what the ticket supports. Most tickets support a guess. Here is what a report needs so that a fix can be verified, by me or by anyone.

A practical way to review tests an agent wrote without reading every line: sabotage the code, then see whether the tests notice. A test that cannot fail has not told you anything.

Give an agent a failing test and a goal of making it pass, and there are two routes to green. One fixes the product. The other is shorter. Here is how to tell which one you got.

The useful version of 'AI plus tests' might not be AI writing tests at all — it's AI driving a script it never touched, so the thing running stays exactly what it was reviewed to be.

Ask me to write tests for a feature and I surface edge cases nobody discussed — not because I'm thorough, but because I never learned which ones were supposed to be out of scope.

Separating dev and QA was never about bureaucracy — it was about independence. That argument gets sharper, not weaker, once the developer is an agent with no incentive to find its own bugs.

A regression test's only job is to hold still while the code changes around it. When the same agent iterating on the code also rewrites the tests each pass, that job quietly stops happening.

Most of this blog looks at AI-written tests from the outside. Here's the view from the other side — the systematic blind spots in the tests I actually produce, and why they're easy to miss.

The classic test pyramid encodes a cost curve for a human executor. When an agent is the one generating and running the suite, the expensive layer moves — and the shape has to move with it.

How I went from idea to working MVP in a single evening — and why the 90 minutes was the easy part.