The idea that artificial intelligence came only to optimize human work may begin to collapse the moment machines stop just suggesting paths and start making their own decisions. This is what happened recently at Amazon Web Services (AWS), which reportedly suffered at least two internal outages carried out by AI agents.
In one case, in December, an AWS system used by customers to analyze costs was down for about 13 hours. According to a report by the Financial Times, agent Kiro, designed to automate workflows, concluded that the best way to solve a problem would be to simply delete and recreate the entire environment. The tool did not suggest the action to an engineer, it performed it itself.
According to employees interviewed by the FT, this was the second recent episode in which AWS AI tools were at the center of service interruptions. In both, engineers allowed the agent to solve the problem without human intervention, something that would normally require validation from more than one person. A senior official described the failures as minor but entirely predictable.
In a statement, Amazon argues that the errors were human, not AI, justifying that any development tool could have caused the same impact. By default, Kiro “requests authorization before taking any action,” Amazon said. However, the engineer involved in the December incident had “broader permissions than expected”, which according to the company is a problem with user access control, not AI autonomy.
One of the problems has been the rush to adopt AI as part of the company’s routine, even in more sensitive functions. Amazon employees told the FT report that they are still skeptical about the usefulness of AI tools for most of their work, due to the risk of errors. However, there is a charge for using these agents.
Amazon has set an internal goal for 80% of its developers to use AI tools to code at least once a week, and is monitoring this adoption closely. Before Kiro, the company was already using Amazon Q Developer, a chatbot to assist engineers in writing code, which would also have participated in one of the outage episodes.
Continues after advertising
After these episodes, Amazon said it implemented several safeguards, including mandatory peer review and employee training. The company also reinforced that the December event had a limited impact, restricted to parts of mainland China, and that the second incident did not directly affect customer-facing services.
In any case, there is a warning for companies that are training their teams to trust agents who do not understand more subjective notions, such as the impacts of decisions on the business. At least for now, human supervision, especially in decision-making, remains essential to avoid errors.
Content produced by .
