Amazon Web Services shares data from internal experiments showing how 'frontier teams' restructure workflows around AI agents to achieve 4.5x to 10x+ productivity gains. Three approaches are described: a pathfinder initiative where six engineers rebuilt the Amazon Bedrock inference engine in 76 days (originally scoped for 30 devs over 12-18 months), a structured 10-day sprint by the Prime Video Financial Systems team achieving ~6x throughput, and in-situ experiments across 50+ Amazon Stores teams using Kiro. Five key practices are identified: investing in agent context (steering files, monorepos, inline docs), accepting an initial slowdown before compounding gains, maintaining parallel agent workloads asynchronously, making intent explicit before coding, and shifting testing left so agents self-correct locally. The core argument is that productivity gains come from restructuring workflows around AI, not just adopting AI tools within existing workflows.
