The Frontier team does more than just use AI to speed up coding. They are redesigning the way software is built. The result is a 4.5x increase in productivity, and in some cases more than 10x.
There are 6 engineers. Seventy-six days. The project took 30 developers 12 to 18 months to complete within one quarter. That’s not a hypothesis. This is what happened when the Amazon Bedrock team stopped treating AI as a coding shortcut and started treating AI as the basis of how work works. The team shipped more product code in five months than in the past 10 years.
The gap between these teams and other teams is rapidly widening. AI coding agents have fundamentally changed the speed at which software is created, but not the speed at which software reaches customers. Commits are exploding and CI/CD pipelines are busier than ever. However, features shipped to production have not kept the same pace. The bottleneck is not the agent’s ability to produce output. It’s about agents having access to the knowledge they need to make good decisions, and teams’ willingness to reshape their work based on that reality.
We call teams that understand this “frontier teams.” They are not limited to elite laboratories. They exist across industries and company sizes and share a common discipline: treating AI adoption as an engineering investment rather than a tool deployment. Any engineering team can be a frontier team. We’ll show you how to get there.
Three paths to AI-native development at Amazon
AI-native software development treats AI as the foundation for building software, with increasingly capable agents directed by human experts. How the team directs the agents determines the outcome. At Amazon, the primary driver for AI in development was to reduce the time developers spend on non-coding tasks such as documentation, coordination, and operations, eliminate technical debt, and minimize coding discrepancies across thousands of small “two-pizza” developer teams. We’ve experimented across hundreds of engineering teams and identified at least three paths. Pathfinder initiatives with experts tackling challenges, structured sprints to execute on a clearly defined plan, and field experiments that split teams in half between existing approaches and AI-adapted workflows. The paths have different structures but converge on the same insight.
of pathfinder initiative It was an experiment of contrasts. Six senior engineers were given one assignment: rebuilding the Amazon Bedrock inference engine. The project was originally estimated to take 30 developers and 12 to 18 months. Rather than increase headcount, the team spent the first few weeks redesigning workflows around AI, moving from individual tasks to goal-driven outcomes, running multiple agents in parallel, and setting up a system where AI could work independently during off-hours. The project was completed in 76 days. Individual developer productivity increased by approximately 20x, as measured by normalized commit velocity (commits per developer per week, adjusting for repository complexity and team size). The number of commits increased from 2 to 40 per week. As measured by lines deployed in production, the team shipped more high-quality code in five months than in any project in the past 10 years.
of structured sprint took a different approach. The Prime Video Financial Systems team conducted a 10-day experiment inspired by the Pathfinder model. Six engineers, one room, no context switching, no on-call duties, no other projects, and limited meetings. Senior engineers spent three weeks upfront breaking down the complexity into broad tasks with detailed requirements. The team used specification-driven development for complex feature work and direct agent-assisted development for tasks where the requirements were already clear. In 10 days, 556 commits were generated compared to 96 in the baseline, reducing the 90-week project estimate to 24 weeks. This equates to nearly 6x throughput and 4x acceleration. They attribute the benefits from AI to a combination of three factors: faster speed at low-judgment tasks (1.5x), ability to focus on high-judgment tasks without context switching (1.5x), and immediate access to domain expertise acquired by the agent (1.5x). If you remove any one element, the gain collapses. The team is currently looking at optimizing these three elements of normal operations using detailed product specifications that encapsulate domain knowledge and autonomous agents that free up focused time.
in Field experimentOf the more than 50 teams surveyed, 25 teams that introduced both new tools and new practices performed better than teams that simply added AI to their existing workflows. Amazon Store used Kiro and purpose-built AI tools to run a structured pilot with a typical development team working on a regular backlog, without special conditions or hand-picked engineers. The median productivity increase was 4.5x, and some teams improved their normalized deployment velocity (features deployed per sprint, normalized to a historical baseline) by more than 10x. Perfect Order Experience now ships features in an afternoon instead of two weeks. WW Grocery reduced design document creation from five days to hours.
The path is different, but the lesson is the same. It’s not just the tools that matter, it’s also the workflow.
5 steps to becoming a frontier team
Across all three paths, the highest performing teams share five practices with common logic. It reduces the barrier to context for agents and increases the surface area of work that agents can perform independently.
This is where the Frontier team deviates from their previous habits. Historical approach optimized for individual code generation speed. The Frontier team optimizes what makes this different: the speed at which accurate, production-ready software reaches customers. This difference drives all of the practices below.
- Invest in the agent’s context. The most advanced teams invest heavily in making projects and knowledge more accessible to agents through agent steering files, team rules, coding standards, testing, and guidance on codebase navigation. The Bedrock infrastructure team placed all code and documentation in a monorepository to preserve inline comments generated by the AI agent, treating them as persistent memory. Teams that skip this step wonder why their agents keep making the same mistakes.
- Slow down to speed up. The above exercises take time and require patience from your team. All of the high-performing teams reported that their work slowed down initially as they learned the model. They encoded cross-functional expertise into reusable steering documents for agents, restructured repositories for LLMs to reason about, added comments and restructured code splitting for AI use. Teams that pushed through that learning curve and defined expected outcomes experienced acceleration for the first time. Teams expecting immediate benefits without changing their workflow were disappointed. Expect to feel slow for the first two weeks. Expect to feel dramatically faster after a few weeks. For a team that quits in Week 2, it can’t get any worse.
- Instead of babysitting your agents, feed them. Frontier teams maintain a steady backlog of broad tasks with clear outcomes, run multiple agents in parallel, and review output asynchronously. Builders report that key features are completed quickly and work is progressing, even when they are not actively waiting for agents to complete tasks. One lead engineer shipped a complete change in just “a few hours of continuous time” as agents worked while engineers moved between code reviews, operational support, and meetings.
- Make your intent clear before you write code. Through structured specifications, detailed requirements documents, or extensive task decomposition, Frontier teams ensure agents have a clear context of what “done” looks like before they start generating code. Some teams using this approach report writing only 1-2% of the code by hand, but the number of commits per person is significantly higher than before.
- “Shift test is over.” The Frontier team builds tools that allow agents to run all integration tests locally and self-fix code before it reaches the pipeline. The Prime Video team invested in automated guardrails, component tests, performance tests, and formatters to catch issues early. Code reviews shifted the focus from code style and naming conventions to interface definitions and architectural decisions.
What technology leaders can do today
Not every team can achieve such results. Teams that skip the context-building phase, treat AI as a drop-in replacement, or expect immediate benefits without reimagining the way they work consistently underperform. Developers across industries are adopting AI coding tools. Not all of them are seeing an increase in production. They’re not using the wrong tools. They are using the right tools within the wrong workflow.
The key points are:
- Transform the way you work so AI can perform at its best.
- Three elements work together to deliver results. AI for low-judgment tasks x uninterrupted focus on high-judgment tasks x instant access to expert knowledge.
- Pilot first, then scale.
The actual starting point is not widespread deployment. A deliberate pilot. Start with a small team willing to spend the first few weeks building the agent context (steering files, spec templates, monorepositories) before writing production code. Empower your team to rebuild workflows. Measure commit speed, deployment frequency, and time to resolution, along with developer satisfaction scores. Then use what you learn to build a playbook for the rest of your organization.
Teams that achieve productivity gains of 4.5x to 10x or more aren’t just adopting better technology. They figured out how to handle it differently. This decision is now available to all engineering organizations. Of course, code commit speed is only part of the story. We want to help you with every aspect of the software development lifecycle, whether it’s streamlining release management, operations, and security operations, or tackling EOL upgrades and the myriad undifferentiated tasks that come with software engineering. Stay tuned for my next blog, where I’ll explain how to approach these.
Learn more about Frontier Team >
To learn more about AI native development, visit AWS Summit New York City.
About the author
Swami Sivasubramanian I’m the Vice President of Agentic AI at Amazon Web Services (AWS). At AWS, Swami has led the development and growth of key AI services such as Amazon DynamoDB, Amazon SageMaker, Amazon Bedrock, and Amazon Q. His team’s mission is to give customers and partners the scale, flexibility, and value they need to confidently innovate with agent AI and build agents that are not only powerful and efficient, but also trustworthy and accountable. Swami also served on the National Artificial Intelligence Advisory Board from May 2022 to May 2025. This advisory committee was tasked with advising the President of the United States and the Office of the National AI Initiative on topics related to the National AI Initiative.
