Amazon’s agentic AI division is building products with six-person teams and shipping them in weeks, bypassing the company’s famous “working backwards” PRFAQ process in favor of building prototypes first. AWS VP Swami Sivasubramanian told GeekWire that AI coding tools have made it faster to build working software than to write the classic six-page Amazon press release document.

The Numbers

The division’s flagship example is Amazon Quick, a desktop app connecting email, calendar, Slack, and documents into a single AI-powered workspace. Sivasubramanian pulled together six engineers in late January 2026. Six weeks later, 200 Amazon employees were using it. Ten weeks in, 10,000 were on it internally. The team shipped publicly on April 28, three months from start, and only wrote the PRFAQ after the product was already in beta. Under the old system, according to GeekWire, the documentation and review alone could have taken as long as building the product.

Other teams followed the same pattern. One team open-sourced Strands, an AWS agent SDK, days after the idea came in a 7 a.m. message to Sivasubramanian. The AI coding tool Kiro was built by a small team using Kiro itself, with one engineer shipping a complex cross-platform notification feature estimated at four weeks in a day and a half.

The Structural Shift

The division is organized into dozens of small teams, many of them two-pizza size (six to eight people). That was Amazon’s founding organizing principle, which much of the 1.5-million-employee company outgrew. AWS CEO Matt Garman carved out agentic AI as its own division last year, and Sivasubramanian deliberately chose small teams. Projects that once required 30 to 40 people, he told GeekWire, can now be done by six to eight.

In a blog post published June 10, Sivasubramanian shared harder numbers. Teams that restructured workflows around AI saw a median 4.5x productivity gain, with some exceeding 10x. One six-engineer team rebuilt the Bedrock inference engine in 76 days, a project originally scoped for 30 developers over 12 to 18 months. That team shipped more production code in five months than in the previous ten years. A Prime Video financial systems team ran a 10-day structured sprint with six engineers and produced 556 commits against a baseline of 96, reducing a 90-week estimate to 24 weeks.

The teams that simply added AI tools to their existing workflows without restructuring did not see comparable results.

Roles Blurring, Costs Shifting

Inside the division, product managers now write code and engineers make product decisions. On the Kiro team, a product manager built the first version of a cost analysis dashboard using Kiro itself.

Sivasubramanian said his division has started tracking AI token spend as an operating cost alongside headcount. The heaviest users consume a few thousand dollars per month. He expects companies will eventually need a full picture of operating expenses that includes the cost of AI agents working alongside employees.

The Broader Pattern

Microsoft’s 2026 Work Trend Index, a survey of 20,000 workers across 10 countries, found that organizational restructuring, not individual AI skill, is the primary lever for AI ROI. OpenAI CTO of applications Vijaye Raji told a recent Technology Alliance event that the company’s top engineers use roughly 100 times more AI tokens than the median, according to GeekWire.

Amazon has cut roughly 30,000 corporate jobs since late 2025 as part of CEO Andy Jassy’s push to reduce bureaucracy. Sivasubramanian framed his division’s approach differently: the same headcount is pursuing a bigger charter, accomplishing in weeks what previously took months.

The Spec Problem

Sivasubramanian learned the limits firsthand. Jet-lagged in India, he used Kiro to rebuild a replication engine he had originally spent four months developing by hand 20 years ago with distinguished engineer Allan Vermeulen. He spent four nights babysitting the agent step by step. On the fifth night, he realized he had not given the agent the right specs and testing environment. Once he did, according to GeekWire, it was done in about two hours.

“The bottleneck is not about the time it takes to build something,” Sivasubramanian told GeekWire. “The bottleneck is about crafting the right specification and the tests and the right product and customer experience.”