AWS キーノートは AI を魔法のように称賛するが、同社のエンジニアたちは異なる現実を語る

INTERVIEW Steve Tarcza, director of Amazon Stores, says his team — StoreGen — exists to help the retail giant's developers move faster and cut friction. But despite the AI mandate, one principle is non-negotiable: nothing ships without a human checking it first. The unit focuses not on AWS customers, but on Amazon's internal development teams for its mammoth retail site and operations. Tarcza spoke to us at the AWS London Summit last week, we met him immediately following the keynote at which Alison Kay, VP and managing director UK and Ireland, told attendees that AI technology feels "like magic." Kay cited as an example how the inference engine behind Bedrock, a generative AI service, was rebuilt in 76 days by six engineers, thanks to using the Kiro agentic coding service. "While the engineers slept, the agents kept building," she said, describing how they "wrote code, tested it, found bugs, fixed them, and deployed it around the clock." The Register suggested to Tarcza that there is some trepidation and worry about this, thanks to well-known security and reliability issues with AI. What issues have his team encountered? "It's the things everybody knows about," he tells us. "It's the hallucinations, it's keeping it within the guardrails." There are cases, he said, where the AI is "even doing work that you didn't ask it to, going further than you wanted to." He is an enthusiast for spec-driven development, which was the key feature of Kiro when it was first previewed in July 2025, the idea being that the AI generates a set of tasks for refinement and approval before writing any code. Does spec-driven development solve problems like hallucination and prompt injection? "No," says Tarcza. "It reduces it at best. And even then, there are cases where it still does go beyond the specification." Kiro was possibly involved in a service outage last year though this was officially denied, with the incident blamed on an employee error. How, then, can agentic AI be made secure and reliable? "We've taken the position that engineers always have to be looking at the output. Nothing ships without someone looking at it and validating it. Spec-driven development helps reduce how much time that takes, because it is then in roughly the form that folks want it to be in," Tarcza tells us. If engineers have to review the code, that implies they have the skills to do so, but with companies – including AWS - busy laying off engineers in the light of AI advances, is it not increasingly likely that code no human has reviewed will go into production? "I take a very strong stance on this," Tarcza says. "If left on its own, that's a natural conclusion to what you may see happen. I think that's a wrong outcome … we can't get to the point where we don't have more junior engineers coming in. We have to continue to grow the talent. We can't end up in a spot where there are not folks to maintain these systems." With all the manual review required, is it right to call AI development magic? "It can be," says Tarcza. "Our engineers spend less than 30 percent of their time working on core engineering, writing code, doing software designs. They spend a lot of time on other aspects of the process. What we've done is remove the friction so they're not writing status reports all the time. It is a magic box in that you can get through these phases faster. "But the idea of it being a magic box that gets you from step one to the final step, it's not there. And I don't think that's the world we want to have." The theme of the keynote was the age of agents but Tarcza is not keen on the term "agentic AI." "I think we should be focusing on taking human-driven processes and re-architecting them with AI at the center," he tells us. Nevertheless, he has the same view regarding agentic actions such as deployment as for AI-generated code. "Right now, every mutating step that an AI might do requires a human to approve it," he says. "That is all the way down to publishing a document for someone to read." He adds that "at least in Stores, we aren't using AI to assist with the deployment. We have great mechanisms that AWS has provided to do automated deployment that is deterministic. If we can have a deterministic system and it accomplishes the outcome that we want, that's preferred." What Tarcza describes is a measured approach that seems in contrast to the breathless agentic AI hype we heard in the preceding keynote. With token costs rising, is AI still worth it? "The highest level thinking about this is, what's the cost of missing a big innovation? The cost of not doing it is almost guaranteed to be higher than the token cost," he says. ®