[Editor’s Note: Agents of Transformation is an independent GeekWire series, underwritten by Accenture, exploring the adoption and impact of AI and agents. See coverage of our related event.]
Amazon is famous for its “work backwards” process. Start with a customer problem, imagine the future where it will be solved, write a press release and FAQ as if it has already happened, focus on the document until it is ready, then go make it a reality.
But last year, it dawned on Swami Sivasubramanian, Amazon Web Services VP of agent AI, that new coding tools made it easier for his teams to build a demo — actual working software — than to write and perfect a six-page Amazon “PRFAQ.”
So they started with a prototype instead.
If something is “a low-risk bet where we want to prove what we think, then I basically say, let’s start building a demo, and then iterate,” Sivasubramanian said in an interview last week, ahead of his keynote speech Wednesday at the AWS Summit in New York.
It is an illustration of how agent tools are reshaping deeply embedded practices and cultures in the workplace. But it’s one of the ways AWS’s AI team is breaking out of the company’s established norms, and in some ways returning to its roots.
Inside Amazon, CEO Andy Jassy says he wants the company to operate as the largest company in the world. The Sivasubramanian classification may be the closest thing to what it looks like in practice.
Back to two pizzas
AWS’s AI agent division is organized into many small teams, many of which are large enough to feed two pizzas. That was the organizational structure that Amazon pioneered in its early days and that most of the company left as it grew to 1.5 million employees.
When Matt Garman, CEO of AWS, rolled out agent AI as part of it last year, Sivasubramanian went with small teams on purpose. It fits the new reality of the AI age: projects that once required 30 to 40 people, he said, can now be done in teams of six to eight.
Example: Amazon’s instant desktop app, which connects to a user’s email, calendar, Slack, documents, and other apps in one workspace, and uses AI to search through them, answer questions, and perform tasks. Amazon’s entry into a market where Anthropic, Microsoft, Google, and OpenAI have caught a lot of attention.
From late January of this year, when Sivasubramanian said it was clear to others in the group that the models below had gotten good enough that the biggest missing ingredient was connecting them to the programs where people actually worked.
He assembled a team of about six engineers to build it. Six weeks later, 200 people within Amazon were using it. Ten weeks ago, there were up to 10,000 inside. The team went back to write the PRFAQ after the product was already in beta, to help prepare their approach to external launch. They were sent on April 28, three months after they started.
Under the old system – writing a PRFAQ, moving it through layers of review – the papers alone could take as long as building and shipping the actual product.
Similar stories play out across the board.
- One team open-sourced Strands, an AWS software development kit for building AI agents, after a member of Sivasubramanian’s team messaged him at 7 a.m. about the idea. After a quick phone call with Garman, they decided to go ahead. In a few days, it was done.
- Kiro, an AI coding tool, was built by a small team on purpose, using Kiro itself to build it. One engineer prototyped a complex cross-platform notification feature for Kiro that took about four weeks of work, and shipped it in a day and a half.
- Amazon’s internal team that rebuilt the identification engine for the company’s Bedrock platform for AI models did it with six engineers in 76 days, a project that was originally expected to take 30 engineers over 12 to 18 months.
Small groups everywhere
What’s happening within Amazon’s AI division is part of a trend across the tech industry toward smaller teams and flat organizations, driven by AI and agents.
Microsoft’s 2026 Work Trend Index, a survey of 20,000 workers in 10 countries, found that the biggest factor behind the real impact of AI on the workplace is not individual skill but whether the organization has restructured to the new technology.
Vijaye Raji, CTO of OpenAI for applications, said during a recent Technology Alliance event that “the company’s ambitions are growing faster than we can hire people” – but the profile of who is being hired is changing. OpenAI is increasingly looking for developers who work with AI tools naturally, and the gap between those who do and those who don’t is stark: OpenAI’s top developers spend nearly 100 times more AI tokens than the median.
All of this leads to a natural question: what does this mean for jobs? Amazon has cut an estimated 30,000 corporate jobs by the end of 2025 as part of what Jassy described as an effort to reduce governance. He said he expects AI to reduce corporate workforces over time.
Similar cuts are playing out across the board, from Meta to Block to LinkedIn, as companies think not only about the roles they need to fill but also how many people they need in total.
Big goals, same team
Sivasubramanian explains the change differently: In his section, the same number of people are now pursuing a large charter. With the new structure, they are able to take on more projects, and faster, accomplishing things in weeks that would have taken much longer in the past.
The nature of roles within those groups is changing, too. Increasingly, product managers write code, and engineers make product decisions. In the Kiro team, for example, the product manager built the first version of a cost analysis dashboard using Kiro itself.
This also requires leaders to work differently. For example, Sivasubramanian said he is careful to monitor which decisions require his approval, even when he is traveling. At the current pace, even four or five days of delay can add about 10% to a team’s shipping timeline.
Managing these groups also raises new questions. Sivasubramanian said his department has begun tracking how much it spends on AI tokens — the basic unit of interaction for the AI model — the way it will track any other operating costs.
So far, the numbers have been manageable: tools like Kiro invest up front in defining the specs and pulling the right core before generating the code, making them more efficient with tokens rather than burning them pointlessly back and forth.
Even the biggest users spend a few thousand dollars a month, he said. But he expects that over time, companies will need a fuller picture of their operating costs that includes not just headcount but the costs of the AI agents working alongside them.
This gets to a larger point: “The bottleneck isn’t about the time it takes to build something,” Sivasubramanian says. “The bottleneck is about building the right specifications and tests and the right product and customer experience.”
In a blog post published last week, Sivasubramanian wrote that teams across the company that redesigned their workflows around AI saw an average 4.5x productivity gain, with more than 10x gains. Teams that simply added AI tools to their workflow did not see the same results.
Coding and testing
That change has created its own challenges. Teams can produce code faster than ever, but if they don’t define what success looks like up front – specs, tests, critical scenarios – agents don’t have much of a chance of success.
Amazon now pushes testing right up to coding time rather than staging, so agents can test their work before anything goes into production.
Sivasubramanian learned this firsthand, the hard way. Earlier this year, exhausted and unable to sleep in his hotel room on a trip to India, he decided to try a fun project: He used Kiro to rebuild a piece of the AWS infrastructure he had developed by hand nearly 20 years ago — a replication engine that still supported key services like S3 and DynamoDB.
He and one of Amazon’s first prominent engineers, Allan Vermeulen, had spent four months making the first purchase. Sivasubramanian thought the agent would make quick work of it. Instead, he spent four nights going back and forth, guarding each step.
On the fifth night, he realized the problem: he hadn’t given the agent the tools to test the output. Once he’s written the right spec and set up the testing environment, it’s done in about two hours. Asked what he did with his rebuilt engine, Sivasubramanian laughed. He didn’t ship it. “Maybe I should do that for you,” she said.
With the right team and a few pizzas, maybe you still can.
