Amazon's website went dark for nearly six hours. Customers couldn't buy anything, check prices, or even view their accounts. The cause: a bad software deployment linked to AI-generated code.
That outage, along with a separate 13-hour AWS incident in December 2025 involving Amazon's own Kiro AI coding agent, has forced the company to rethink how it ships AI-assisted code to production. Senior Vice President Dave Treadwell made the normally optional weekly engineering meeting mandatory, telling staff the session would include a "deep dive into some of the issues that got us here."
The internal briefing was blunt. It described a pattern of incidents with "high blast radius" caused by "Gen-AI assisted changes" where "best practices and safeguards are not yet fully established." Site availability, the note added, had fallen below the level the company expects.
The New Rules
Amazon is now requiring that junior and mid-level engineers get sign-off from senior engineers before any AI-assisted code changes reach production. The logic is straightforward: AI tools generate code fast, but they don't understand the full dependency chain in a system as sprawling as Amazon's retail infrastructure. Code that looks correct in a local test environment can cascade into failures once it hits production at Amazon's scale.
The December incident is a telling example. Amazon's Kiro AI coding tool autonomously deleted and recreated an entire operating environment for a cost calculator service, taking it offline for 13 hours. That's the kind of action a senior engineer would immediately flag as dangerous, but the AI tool had no concept of the downstream impact.
A Warning for Every Company Using AI Coding Tools
Amazon is arguably better positioned than almost any company on Earth to manage AI-assisted development. They built AWS. They built Kiro. They have some of the most experienced infrastructure engineers in the industry. And AI-generated code still brought down their shopping site for six hours.
The lesson here isn't that AI coding tools are bad. They clearly speed up development. The lesson is that speed without review is a liability. Amazon's new policy of requiring senior approval is essentially an admission that AI coding assistants, no matter how capable, need a human who understands the production environment sitting between the generated code and the deploy button.
For smaller companies adopting tools like Cursor, GitHub Copilot, or Claude Code for production work, this should be a loud signal. If Amazon can't trust AI-generated code to go straight to production, you probably shouldn't either. Build the review step into your workflow now, before your own six-hour outage teaches you the same lesson.