The team that stopped writing tests

At Meta I worked on notifications, a team of about twenty. We shipped fast and the systems were huge - hundreds of millions of notifications a day - but our test coverage sat at about 35%. Nobody was proud of it. Everyone knew tests mattered. There was just always something more urgent, and writing tests for someone else's code is the task that quietly falls off every sprint.

So we tried something that felt a little crazy at the time: we let two LLM agents do the boring part. One agent wrote the tests. A second, separate agent reviewed them - edge cases, assertions that actually assert something, none of the confident nonsense a single model will happily hand you. A human made the final call on what got merged.

Coverage went from about 35% to up 17 points - not because the models were magic, but because the review loop made their output trustworthy enough to merge.

That is the lesson this whole guide is built on. The models were not the hard part. Trust was. A single agent producing tests nobody reviewed would have been rejected within a week, and honestly it would have deserved it. Two agents checking each other, with an engineer holding the merge button, produced work the team actually relied on.

I have since built and reviewed AI tooling in a very different world - a live risk platform at a bank, where a wrong number is not an embarrassing bug, it is an incident. The bar for trusting AI-generated work is higher there, and it should be. But the pattern is exactly the same: generation is cheap, verification is the product.

Do not ask "what can the model do?" Ask "what would make my team trust its output enough to ship it?" Build that, and adoption stops being a fight.

This is the first chapter of a short field guide I wrote: Shipping AI Dev Tools Teams Actually Trust. It covers picking the first AI feature a team will keep, the writer-reviewer pattern, evals before vibes, blast radius, rollout that sticks, production monitoring, and the cadence that ties it together. It is $34 here: https://ishwarjha4.gumroad.com/l/shipping-ai-dev-tools. If this chapter was useful, the rest goes deeper.