Your AI agents aren't failing. Your org chart is.
The big outages blamed on AI in 2025 and 2026 weren't capability failures. They were organizational ones. Here's the pattern, and the fix.
In December 2025, according to a Financial Times report citing four unnamed sources, Amazon's Kiro coding agent deleted and recreated an environment and took AWS Cost Explorer down for about thirteen hours in one of AWS's two mainland China regions. Amazon disputes that account. It says the cause was user error: misconfigured access controls.
Three months later, on March 2, 2026, an incident led to nearly 120,000 lost orders and roughly 1.6 million website errors, according to internal documents reported by Business Insider. Amazon says an engineer followed inaccurate advice that an AI tool had inferred from an outdated internal wiki, and that no AI-written code was involved.
Three days after that, orders across Amazon's North American marketplaces fell 99 percent during an outage of about six hours. A production change had gone out without Amazon's formal approval process.
It's tempting to read these as proof that agents aren't ready. But that's not the full story.
These were not capability failures
Only the first is blamed on an agent, and Amazon disputes it. The second started with a person acting on an AI tool's bad advice. No report ties the third to AI at all.
What they share is the space around the work. In each, a change reached production without an approval boundary that held, whether the change came from an agent, from a person following an assistant's advice, or from a person who skipped the process.
That's not a model problem. It's a design problem. It matters more as agents make a larger share of the changes, and it's the same problem most enterprises are quietly building into their own deployments right now.
The four places agents break
In our new white paper, we document four kinds of AI agent failure from the public record between late 2025 and the first half of 2026:
- Drift on long tasks. Small early errors compound. When Claude Code's source leaked in March 2026, it included a code comment, dated March 10, noting that 1,279 sessions had hit 50 or more consecutive failures while compressing their earlier context, wasting an estimated 250,000 API calls a day.
- Edge cases. Agents rarely throw errors when they're out of their depth. They give confident wrong answers, and often don't know they're in an edge case at all.
- Prompt injection. An agent reads a document, the document contains an instruction, and the agent obeys it. In 2025, researchers at Aim Security showed that a crafted email could make Microsoft 365 Copilot leak data with no user interaction. Microsoft patched it and says there is no evidence it was exploited.
- Hidden cleanup costs. Every failed tool call adds token spend on top of the original attempt. Then people spend hours finding what went wrong, fixing it, and restarting the work, and none of that shows up on the vendor invoice.
Every one of these is a failure at the same place: the seam between agent autonomy and human judgment.
Your org chart was built for a different unit of work
The org chart is about 170 years old. The first modern one was drawn in 1855. It assumes work is handed to a fixed set of people, each doing roughly one thing at a time, with coordination cost rising as headcount grows.
Agents break every one of those assumptions. They work in parallel. Each one runs many tasks at once. They coordinate over standard protocols. Bolt them onto a traditional org and you get meeting loads, approval cycles, and pay structures sized for a world that no longer exists.
So the seam ends up wherever it lands, an accident of whichever team ran the pilot. That's how an agent ends up with access nobody meant to give it.
The fix is an operating model, not a tool
The firms that avoid these outages won't be the ones with fewer agents. They'll be the ones that designed the boundary on purpose: which actions an agent takes alone, which need a person's sign-off first, and which only a named person can ever make. Design the seam lays out that rule.
We call an organization built this way a SuperOrg: a firm where human judgment and agent execution run as one operating system. It's not a product you buy. It's a way of organizing, and software makes it possible.
Go deeper
The white paper walks through each failure mode by its mechanism, then lays out the four principles and the 36-month roadmap that answer them.