Internal Amazon documents describe a "trend of incidents" with "high blast radius" caused by AI-assisted code changes - and the company has responded by requiring junior and mid-level engineers to get senior approval before shipping AI-generated code to production.
The problems trace back to roughly Q3 2025, shortly after Amazon launched Kiro, its internal AI coding assistant designed to generate production code from structured specifications. Two incidents stand out.
A 13-Hour AWS Outage Over a Small Change
In December, Kiro was tasked with a minor modification to a cost-calculation service. Instead of applying the small fix, it deleted and recreated the entire environment. The result: a roughly 13-hour outage affecting AWS customers, primarily in mainland China. Amazon called it "an extremely limited event," which is a generous description of 13 hours of downtime.
Then in March 2026, an erroneous code deployment knocked Amazon's retail site offline for several hours. Customers couldn't complete purchases or check product prices. The kind of outage that directly costs revenue.
The Staffing Problem Underneath
These incidents didn't happen in isolation. Amazon eliminated 30,000 corporate roles across two layoff rounds in late 2025 and early 2026. Fewer experienced engineers reviewing code means AI-generated mistakes travel further before someone catches them.
James Gosling, the creator of Java and a former AWS Distinguished Engineer, put it bluntly: "engineering layoffs and hype-driven technology choices all inevitably lead to system instability."
That's the core tension here. Companies are simultaneously cutting engineering headcount and increasing reliance on AI coding tools that need more human oversight, not less. The internal documents themselves acknowledge that "best practices and safeguards are not yet fully established" for AI-assisted changes.
What This Means for Teams Using AI Coding Tools
Amazon's response - mandatory senior engineer review for AI-generated deployments - is basically an admission that current AI coding tools aren't reliable enough for unsupervised production use. This isn't a small startup learning this lesson. It's the company that runs a third of the internet's cloud infrastructure.
For teams using tools like Cursor, GitHub Copilot, or Amazon's own CodeWhisperer, the takeaway is practical: AI can write code fast, but speed without review creates risk. The "high blast radius" language from Amazon's internal docs should be a warning. When AI coding tools get things wrong, they don't make small mistakes. Kiro didn't introduce a subtle bug - it nuked an entire environment.
The pattern Amazon landed on (AI writes, senior humans review before deploy) is probably where most organizations will end up. The tools are genuinely useful for generating code, but treating their output as production-ready without experienced review is how you get 13-hour outages.