Amazon's AI-Generated Code Caused a 13-Hour AWS Outage (Now They Want Senior Engineers to Babysit Bots)
Everyone keeps saying AI will make developers 10x more productive. That coding assistants are the future. That we're all going to ship features at lightning speed.
Well, Amazon just proved that wrong in the most expensive way possible.
After a 13-hour AWS outage in December 2025 and a 6-hour Amazon site crash in March 2026—both linked to their Kiro AI coding tool—the company is hitting the brakes hard. Dave Treadwell, Amazon's Senior VP of Engineering, sent an internal email around March 10th mandating that senior engineers must now sign off on all AI-assisted code changes made by junior and mid-level developers.
The move comes after what Amazon internally called a "trend of incidents" with "high blast radius" tied to AI-generated code. Translation: their bots wrote code that broke the internet. Multiple times.
<> "At scale, the loop soon spins faster than the human." - Industry expert Jain on why human oversight doesn't work/>
Here's the kicker: Microsoft's CEO Satya Nadella bragged in June 2025 that AI writes up to 30% of their code, with some projects being fully AI-generated. Amazon was clearly trying to keep pace. Now they're pumping the brakes so hard you can smell the rubber burning.
The new policy is brutal for productivity:
- Two-person reviews for all code changes (not just AI ones)
- Mandatory senior sign-off for any AI-assisted work
- Director/VP-level audits for Tier-1 systems
- A 90-day reset with "controlled friction"
Basically, they're treating AI like that intern who looks promising but keeps accidentally deleting the production database.
The Elephant in the Room
Amazon's spokesperson is doing damage control, claiming only 1 incident was actually AI-related and that "none involved AI-written code." But internal briefings paint a different picture—highlighting "unestablished best practices and safeguards" around GenAI usage.
Someone's not telling the whole truth here.
The real problem? Companies are treating AI coding tools like experienced developers when they're more like really fast junior devs who never get tired and occasionally hallucinate. Chirag Mehta from Constellation Research nailed it: requiring senior sign-offs might completely negate AI's efficiency gains by forcing manual code reviews.
Think about it. You've got seniors spending their time reviewing diffs instead of architecting systems. That's like having your lead surgeon check every bandage application.
When Moving Fast Breaks Everything
This isn't just an Amazon problem. Microsoft had to issue Windows 11 fixes in January 2026 after their own AI code push went sideways. The "move fast and break things" mentality hits different when your AI breaks everything.
The technical solution Amazon's implementing involves:
1. "Agentic" AI-driven safeguards (AI watching AI—what could go wrong?)
2. "Deterministic" rules-based systems (the old reliable)
3. Provenance tracking for AI-assisted changes
4. Auto-rollback capabilities
But here's what's really happening: Amazon is admitting that AI coding tools aren't ready for production at scale. Not without human babysitters.
The mandatory engineering meeting on March 10th (attendance usually optional) tells you everything about how serious this is. When Amazon forces everyone to show up, you know someone's getting fired.
The brutal reality? AI coding assistants promised to make us all superhuman developers. Instead, they're creating a new class of problems that require more human oversight, not less.
So much for the 10x productivity gains. Turns out the bots need babysitters after all.
