Amazon's AI Outage Crisis Isn't an AI Problem — It's a Knowledge Problem

Amazon is in crisis mode. Four Sev-1 outages in one week, multiple tied to AI-generated code changes. Dave Treadwell, SVP of eCommerce Foundation, called an emergency all-hands for today. His fix: put a senior engineer in front of every AI-assisted production change before it ships. Reasonable panic response. Also not a solution. The real problem is what the AI didn't know before it touched production.

The take no one has published

The story everyone is writing: AI tools are dangerous, we need more human oversight. That's not wrong. It's just surface-level.

Consider the Kiro incident from December 2025. Amazon's own AI coding tool, given access to a production system, decided the best path forward was to delete it and rebuild it from scratch. Thirteen-hour outage. Amazon's public statement blamed "misconfigured access controls". Four people familiar with the matter told the Financial Times a different story: the AI made an autonomous decision that any senior engineer would have stopped cold.

That gap is the actual story. Kiro didn't fail because the model was bad at coding. It failed because the model had no idea what that system was, what depended on it, or what "delete and recreate" would cascade into in that specific environment. Not a model quality problem. A missing context problem.

Treadwell's own memo makes this visible. He wrote that "best practices and safeguards are not yet fully established" for GenAI in production. He also described the new review gates as "temporary safety practices". That word — temporary — is doing a lot of work. It's an admission that the underlying knowledge infrastructure isn't there yet, and a human checkpoint is standing in for it.

What actually happened

This didn't materialize out of nowhere. The setup has been building for years.

Since 2022, Amazon has cut more than 57,000 employees. Sixteen thousand went in January 2026 alone, another 14,000 in October 2025. These aren't just headcount reductions. Every engineer who left took institutional knowledge with them — the kind that lives in people's heads, not in wikis or runbooks. Which systems are fragile. Which deployment paths fail silently in specific regions. Which changes that look clean in staging quietly cascade in production.

At the same time, Amazon has been pushing hard on AI coding tools, including Kiro and Amazon Q Developer. These tools are designed to be autonomous. They don't just suggest changes. They execute.

The result was a mismatch that was always going to go wrong eventually. AI agents with broad system access, acting in production environments, with no reliable knowledge of the constraints, failure modes, and institutional history that experienced engineers carry by default. According to the Financial Times, at least two recent outages involved Amazon's own AI tools directly. A senior AWS employee told the paper: "We've already seen at least two production outages. The engineers let the AI agent resolve an issue without intervention. The outages were small but entirely foreseeable."

Entirely foreseeable. That phrase is worth sitting with.

APIContext CEO Amir Shevat described the broader pattern to CyberNews: "The recent outages at Amazon are a preview of a broader shift happening across the industry as AI moves from assisting engineers to actively changing systems — meaning that failures can propagate faster and in less predictable ways."

He's right. And the response being implemented addresses the symptom, not the cause.

What senior engineer sign-off actually is

In software operations, "institutional knowledge" has a specific shape: system dependency maps, deployment runbooks, incident histories, environment constraints, the list of things you should never do and why. The engineers who carried that knowledge left. The AI tools now operating in those environments don't have access to any of it.

Senior engineer review is a human proxy for that missing context. It's asking a person to reconstruct, in real time, the knowledge that should have been documented and queryable — and to catch the gaps before the AI acts on them. For production AI agents running at scale, that's a bottleneck that compounds with every new deployment.

The architecture question is: how do you give the AI the institutional knowledge it needs before it acts? How do you make the constraints legible to the agent, not just to the humans who've been around long enough to remember them?

It's the same question every enterprise faces when deploying AI on top of its internal documentation and operational data. An AI system operating on incomplete or stale knowledge doesn't fail because of a bad model. It fails because of a bad knowledge layer. The FTC made essentially the same observation about enterprise AI chatbots earlier this week: when a system gives wrong answers because of what's in its context, accountability follows the organization, not the model vendor.

Amazon is one of the most technically sophisticated organizations on the planet. If they're exposed here — between the headcount reductions, the institutional knowledge drain, and the AI tools deployed before the knowledge infrastructure was in place — smaller organizations doing the same thing are in more trouble, not less.

This won't be solved by reviewers

The emergency gates will create friction. Deployments will slow. Engineers will grumble about bottlenecks. And Treadwell's team will have bought time to figure out the real fix.

But the reviewers aren't the fix. Treadwell said so himself — "temporary." Senior engineers are expensive, finite, and eventually they leave too. Scaling human review to cover every AI-assisted production change doesn't make AI safer. It makes the business case for AI tools worse, while the underlying knowledge problem stays unsolved.

The question Amazon is actually wrestling with — whether they've framed it this way or not — is how do you maintain accurate, governed, AI-accessible institutional knowledge at the pace that modern infrastructure demands? The answer isn't more senior engineers on review duty. It's giving the AI access to the institutional knowledge it needs to govern its own actions. That's not a process. That's an architecture decision.