If you’ve ever shipped code on a Friday and immediately regretted it, imagine doing that with AI-generated code… across one of the largest e-commerce platforms on Earth.
Amazon just held a mandatory engineering meeting after a string of outages hit its retail website and app, including a six-hour crash last week that left customers unable to check out, see prices, or access their accounts. An internal briefing note described the incidents as having a “high blast radius” and being related to “Gen-AI assisted changes.”
Here’s what we know
- Senior VP Dave Treadwell acknowledged that “best practices and safeguards” around AI coding tools aren’t fully established yet
- Junior and mid-level engineers now need senior sign-off on any AI-assisted code changes
- AWS separately suffered a 13-hour outage in December after its Kiro AI tool deleted and recreated an entire coding environment
- Amazon has dashboards tracking whether engineers hit minimum daily AI usage targets
- Amazon disputes that AI wrote the bad code; they say it’s a “user error” and “protocol” problem
Meanwhile, on Reddit, current and former Amazon engineers paint a grimmer picture. One described “on-calls using AIs to fight each other’s AIs in a proxy war of blame.” Another said delivering projects matters more than whether projects actually work. Sounds healthy.
Amazon also recently laid off 16,000 workers in January, mandated 80% AI tool usage targets (again, dumb; why is “usage” the metric and not “make our products better?”). Oh, and they’re spending $200B on capex this year. So do the math and we have fewer engineers + more AI-generated code = more mandatory WTF did we just break meetings.
Why this matters
As “The Primeagen” and AI researcher Demetri Spanos discussed this week: the models are smart, but they’re not THAT smart. Actually, the practices around how they are used are the real problem.
Most of the excitement since December 2025 has come from the maturation of agent loops and team workflows, not from raw model improvements alone. AI can write code, but the code it writes is often much more verbose than it needs to be, and when some poor human (likely you) has to go in and read it, unlike other large, well-known codebases around the world, no one knows wtf this one says… including you.
So, how do you set up your processes to catch the code that AI writes incorrectly? And how do you set up your systems to get AI to write code efficiently? As Dylan Patel said in a recent Matt Berman interview, some of his team is spending something like $5K a day on Claude Code tokens. How many tokens do you think that guy is spitting out?
Amazon will figure this out. But they’re learning the lesson every company using AI for production code will eventually learn: the speed you gain means nothing if you can’t trust the output.
Editor’s note: This content originally ran in the newsletter of our sister publication, The Neuron. To read more from The Neuron, sign up for its newsletter here.