Blog
Notes on shipping fast without breaking things.

Your devs are 4x faster. Your company isn't.
Apiiro studied the codebases of large enterprises. Developers using AI assistants produced three to four times more commits. Over six months, the number of security findings from AI-generated code rose tenfold.
Testing companies are starting to admit their own problems
Start with mabl. They had 288 tests running on every code change, and they cut the gate to 17. The reasoning is arithmetic. Even if every test works 99% of the time, the chance that all 288 pass together is about 5%. The alarm went off so often that nobody believed it anymore.
Who reads your code first, you or them?
One wrong character gave total control of the machine. It was a single mistyped letter in plain sight, in software that runs inside your phone and half the internet. It stayed hidden for years. Then an AI agent read the entire codebase and found it in a single pass.

Checking became the step you could skip
Amazon lost about 6.3 million orders because of code that no human reviewed. The cause was an AI-assisted change pushed straight to production, and the result was a six-hour outage of the storefront. On 5 March 2026 U.S. order volume dropped by 99%, roughly 6.3 million orders (Financial Times and CNBC, from Amazon's internal documents).

Your tests were written by the thing they're testing
Your tests were written by the thing they are testing. This is rarely said about AI-generated tests.

Convincing is what slips through review
The best code shipped all quarter took down the checkout page at 2am. It had passed pull request review in twenty minutes, with no comments on it.

Zero percent said very confident
A 2026 survey asked engineering leaders how confident they were that AI-generated code would work in production. Zero percent said they were "very confident." Not a small number. None.