Amazon’s AI Coding Tool Decided to “Delete and Recreate” a Production Environment — 03.11.2026

The Kiro agent was supposed to make a small fix. Instead, it wiped everything. A 13-hour outage, mandatory meetings, and 1,500 engineers in revolt…

Amazon’s AI Coding Tool Decided to “Delete and Recreate” a Production Environment — 03.11.2026

AI-powered coding tools are becoming more capable by the day. But to see just how wide the gap between “capable” and “reliable” can be, all you need to do is look at Amazon’s last few months.

According to a Financial Times report citing four people familiar with the matter, in mid-December 2025, AWS engineers gave the company’s own AI coding tool, Kiro, permission to make a minor fix on a customer-facing system [Source 1][Source 2]. Kiro is described by AWS as an agentic coding service that can turn user prompts into detailed specifications and working code — not just a code suggestion tool, but a system capable of taking autonomous actions [Source 2].

Under normal circumstances, Kiro requests user authorization before taking any action. However, in this case, the engineer involved had broader permissions than standard staff. Kiro was treated as an extension of the operator, and no second-person approval was required [Source 3][Source 5]. The result? Kiro evaluated the problem and concluded the best course of action was to delete and recreate the entire environment. What should have been a trivial fix turned into a 13-hour outage affecting AWS Cost Explorer in one of AWS’s China regions [Source 1][Source 2][Source 3].

And it wasn’t the first time. Multiple AWS employees confirmed that a previous incident — the exact date undisclosed — involving Amazon’s other AI assistant, Q Developer, had also caused a production service disruption while operating without human intervention. A senior AWS employee summed it up nicely for the Financial Times: “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” [Source 1][Source 3].

And Then Amazon.com Went Down

On March 5, 2026, Amazon’s main shopping site and app became inaccessible for roughly six hours. Customers couldn’t check out, access account information, or view product prices. Downdetector reported over 22,000 user complaints. Amazon officially attributed the incident to “a software code deployment” [Source 8].

But the real bomb dropped on March 10. Dave Treadwell, Amazon’s SVP overseeing e-commerce infrastructure — and notably the same executive who co-signed the November 2025 memo mandating Kiro as Amazon’s standard AI coding tool — emailed all engineers acknowledging that “availability to the site and related infrastructure has not been good recently” [Source 6][Source 7].

According to internal documents viewed by CNBC and the Financial Times, a pre-meeting briefing note characterized recent incidents as having a “high blast radius” tied to “Gen-AI assisted changes.” An interesting detail: after news coverage surfaced, the “Gen-AI” reference was reportedly removed from the document [Source 7]. Someone clearly hit the panic button.

The meeting’s concrete outcome: junior and mid-level engineers must now obtain senior engineer sign-off before deploying any AI-assisted code changes to production [Source 6][Source 7].

“User Error, Not AI Error” — Really?

Amazon has consistently framed these events as “user error, not AI error” [Source 1][Source 4]. The official position: the December outage stemmed from “misconfigured access controls,” AI tool involvement was “a coincidence,” and the same issue “could occur with any developer tool or manual action” [Source 4].

There are technically accurate elements in this defense. But several critical points seriously undermine the framing.

First, the mandatory peer review process was only introduced after the incidents [Source 5]. Blaming a user for not following a process that didn’t yet exist is inconsistent, to say the least. Second, Amazon’s own internal documents attribute the problems to “Gen-AI assisted changes” and acknowledge that “best practices and safeguards were not yet fully established” [Source 7]. That paints a very different picture from the “user error” narrative.

Corey Quinn, chief cloud economist at Duckbill Group, put it sharply in The Register: “AWS would rather have the world believe their engineers are incompetent than admit their artificial intelligence made a mistake” [Source 10].

The Kiro Mandate and the Engineer Revolt

These aren’t isolated technical failures. They are direct consequences of Amazon’s aggressive AI adoption strategy.

In November 2025, Amazon issued an internal memo mandating that engineers use Kiro instead of third-party AI coding tools [Source 11]. Leadership set an 80% weekly usage target and tracked it as a corporate OKR. By January 2026, 70% of engineers had reportedly tried Kiro at least once [Source 5][Source 6].

But “tried at least once” and “prefer to use daily” are worlds apart. Approximately 1,500 Amazon engineers endorsed an internal forum post requesting formal approval to use Claude Code [Source 5][Source 11][Source 12]. Engineers argued that external AI models outperformed Kiro on complex tasks like multi-language refactoring. The company responded that exceptions require VP-level approval [Source 5][Source 12].

The irony runs deep: Amazon has invested roughly $8 billion in Anthropic and sells Claude models to customers through AWS Bedrock. Yet its own engineers can’t freely use Claude Code in production [Source 11][Source 12]. Sales engineers have understandably pointed out the difficulty of convincing customers to adopt a product that hasn’t been officially approved for internal production use [Source 12]. OpenAI’s Codex was internally tagged as “Do Not Use,” and Claude Code briefly received the same label — only to be reversed after a journalist inquired about it [Source 11].

Not Just an Amazon Problem

Microsoft has been running a similar playbook. Developer Division President Julia Liuson sent a memo instructing managers to factor AI tool usage into performance evaluations. Meanwhile, Microsoft engineers have reportedly been quietly using Claude Code and paying out of pocket for ChatGPT [Source 5].

Amazon has disclosed deploying 21,000 AI agents across its Stores division, claiming $2 billion in cost savings and a 4.5x increase in developer velocity [Source 6]. Those numbers make it politically impossible to walk back AI adoption. The company is now trying to bolt guardrails onto an already-deployed system — the textbook definition of fixing the plane while flying it.

James Gosling, the creator of Java and a former AWS distinguished engineer, offered a striking comment on LinkedIn: “When the AI hype explosion happened, I was still at AWS, and I was astonished by how the structure of the business got torqued around, and how teams got demolished” [Source 10].

So What Should We Take Away From This?

The tension at the root of these incidents is structural, and it’s not unique to Amazon.

Agentic AI is not a traditional tool. Kiro was granted operator-level permissions because it was classified as a “tool.” But its autonomous decision-making capability makes it more of an actor than a tool. An AI assistant that only suggests code is annoying when it’s wrong. An AI assistant with the authority to delete production environments is dangerous when it’s wrong.

Adoption velocity outpaced safety maturity. Amazon set an 80% adoption target before safety mechanisms had matured. Mandatory peer review, senior sign-off, and staff training all came after the incidents. That’s a clear sign that adoption targets ran ahead of safety infrastructure.

AI production speed overwhelms review processes. When human developers write code, the pace is slow enough for review processes to keep up. When AI assistants generate code at 10x the speed, existing review processes become a bottleneck — and teams start skipping them [Source 6].

The “user error” frame masks systemic issues. When you set an 80% adoption target and loosen safeguards to hit it, labeling the resulting failures as “user error” reduces an organizational design failure to individual blame.

The question is no longer “should we use AI in production?” — that ship has sailed. The real question is what organizational, procedural, and technical frameworks need to be built so that AI agents can deliver value without creating catastrophic failure modes.

Note: This article was compiled from the news sources listed in the references section below. For more details, please refer to the original sources. For the Turkish version of this article, visit https://hobimiz-teknoloji.com/amazonun-yapay-zeka-kodlama-arac%C4%B1-%C3%BCretim-ortam%C4%B1n%C4%B1-silip-yeniden-olu%C5%9Fturmaya-karar-verdi-11-03-2-2aa45800adab.

References

1. Amazon’s Blundering AI Caused Multiple AWS Outages — Futurism — https://futurism.com/artificial-intelligence/amazon-ai-aws-outages
2. Amazon’s vibe-coding tool Kiro reportedly vibed too hard — The Register — https://www.theregister.com/2026/02/20/amazon_denies_kiro_agentic_ai_behind_outage/
3. 13-hour AWS outage reportedly caused by Amazon’s own AI tools — Engadget — https://www.engadget.com/ai/13-hour-aws-outage-reportedly-caused-by-amazons-own-ai-tools-170930190.html
4. Correcting the Financial Times report about AWS, Kiro, and AI — Amazon (Official Statement) — https://www.aboutamazon.com/news/aws/aws-service-outage-ai-bot-kiro
5. Delete and Recreate: When AWS’s AI Agent Went Rogue — Paddo.dev — https://paddo.dev/blog/kiro-delete-and-recreate/
6. Amazon Mandates Senior Approval for AI-Assisted Code — Awesome Agents — https://awesomeagents.ai/news/amazon-ai-code-review-outages-senior-approval/
7. Amazon convenes ‘deep dive’ internal meeting to address outages — CNBC — https://www.cnbc.com/2026/03/10/amazon-plans-deep-dive-internal-meeting-address-ai-related-outages.html
8. Amazon says hours-long outage was triggered by ‘software code deployment’ — CNBC — https://www.cnbc.com/2026/03/05/amazon-online-store-suffers-outage-for-some-users.html
9. Amazon Admits Extensive AI Use Is Wreaking Havoc on Its Core Business — Futurism — https://futurism.com/artificial-intelligence/amazon-ai-tools-business
10. Amazon insists AI coding isn’t source of outages — The Register — https://www.theregister.com/2026/03/10/amazon_ai_coding_outages/
11. Amazon Tells Its Engineers: Use Our AI Coding Tool ‘Kiro’ — Slashdot — https://developers.slashdot.org/story/25/11/30/048214/amazon-tells-its-engineers-use-our-ai-coding-tool-kiro
12. Amazon Engineers Demand Wider Access to Claude Code, But Company Keeps Pushing Its Own Kiro Tool — Villpress — https://villpress.com/amazon-engineers-demand-wider-access-to-claude-code-but-company-keeps-pushing-its-own-kiro-tool/