Amazon Teams Hit 10x Productivity With AI-Native Workflows

Engineering teams are rewriting the rules

A six-person Amazon Bedrock team recently delivered a project originally scoped for 30 developers over 12 to 18 months. They finished in 76 days. Individual developer productivity jumped roughly 20-fold, measured by normalized commit velocity—from two commits per week to 40. The team shipped more production code in five months than in the previous decade.

The breakthrough wasn't a better AI coding tool. It was a complete redesign of how the team worked, according to details first reported by AWS.

Across hundreds of Amazon engineering teams, a pattern has emerged: AI coding agents have dramatically increased the volume of code written, but not necessarily the rate at which features reach customers. Commits are surging while production deployments lag. The constraint isn't the agent's output—it's the agent's access to context and teams' willingness to restructure workflows around that reality.

Why it matters

Most organizations treat AI coding assistants as productivity boosters within existing processes. Amazon's data suggests that approach leaves most of the value on the table. Teams that restructured workflows around AI capabilities achieved median productivity gains of 4.5x, with some exceeding 10x. The gap between these "frontier teams" and everyone else is widening, and it's driven by process redesign, not tool selection.

Three experimental paths, one insight

Amazon tested AI-native development through three distinct approaches. The pathfinder initiative with the Bedrock team focused on goal-driven outcomes rather than discrete tasks, running multiple agents in parallel and enabling AI to work independently during off-hours.

A structured sprint by the Prime Video Financial Systems team compressed a 90-week project estimate to 24 weeks. Six engineers in a 10-day focused session produced 556 commits against a baseline of 96. They attributed gains to three multiplying factors: 1.5x acceleration on low-judgment work, 1.5x higher focus without context-switching, and 1.5x from instant access to agent-captured domain expertise. Remove any single factor and the gains collapse.

In broader pilots across 50-plus Amazon Stores teams, the 25 teams that implemented both new tools and new practices outperformed those that simply added AI to existing workflows. The median gain was 4.5x in normalized deployment velocity. One team now ships features in an afternoon instead of two weeks.

Five practices that separate frontier teams

The highest-performing teams converge on five disciplines:

Invest in agent context. Top teams create agent steering files, encode team conventions and coding standards, and restructure repositories so language models can reason over them effectively. The Bedrock team placed all code and documentation into a monorepo and preserved inline commentary that AI agents generated, treating it as persistent memory.

Slow down to speed up. Every high-performing team reported an initial slowdown as they encoded cross-functional expertise into reusable guidance for agents. Teams that pushed through the first two weeks of learning curve saw compounding acceleration afterward.

Feed agents instead of babysitting them. Frontier teams maintain backlogs of well-scoped tasks with clear outcomes, run multiple agents in parallel, and review output asynchronously. One principal engineer shipped a complete feature with only a couple hours of contiguous time because the agent worked during meetings and code reviews.

Make intent explicit before code gets written. Whether through structured specifications or detailed requirements documents, teams ensure agents understand what "done" looks like. Some teams report handwriting only 1–2% of their code while pushing significantly more commits per person.

Shift testing left. Teams build tooling so agents can run integration tests locally and self-correct before code reaches the pipeline, moving code reviews toward interface definitions and architectural decisions rather than style and naming.

The starting point

The practical path forward isn't a broad rollout. Amazon recommends starting with a small team willing to spend initial weeks building agent context before writing production code. Give them a mandate to restructure workflows, measure commit velocity and deployment frequency alongside developer satisfaction, then use their learnings to build a playbook for the organization.

These details were first reported by AWS Vice President for Agentic AI Swami Sivasubramanian on the AWS Machine Learning Blog.