Sign InOpen Brain
AI EngineerVideoSource Linked

From AI-Assisted to AI-Native: Building a Frontier Development Team — Clare Liguori, AWS

Amazon’s pilot suggests coding-agent gains come from redesigning the workflow: prepare agent-readable context, fast local feedback, scoped specs, and parallel work instead of constant chat.

AI Engineer · Aug 28, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Amazon observed **50 teams** working in existing codebases for most of a year, measuring deployment velocity rather than commits. Half gained under 3x; the stronger cohort reached a **4.5x median**, with some exceeding **10x**. Tool choice was not the main separator because 90% used Kiro.

Practical Implication

Treat agent adoption as engineering work. Improve error messages, tests, local mocks, context, task specifications, and tools so agents can run longer and self-correct. Keep steering files lean as models improve, and feed multiple agents scoped work instead of supervising one continuous chat.

Agent-Ready Context
Amazon observed **50 teams** working in existing codebases for most of a year, measuring deployment velocity rather than commits. Half gained under 3x; the stronger cohort reached a **4.5x median**, with some exceeding **10x**. Tool choice was not the main separator because 90% used Kiro.

Treat agent adoption as engineering work. Improve error messages, tests, local mocks, context, task specifications, and tools so agents can run longer and self-correct. Keep steering files lean as models improve, and feed multiple agents scoped work instead of supervising one continuous chat.

The figures come from internal Amazon pilots, including unusually structured experiments, so they may not transfer directly. Teams initially slowed down, and longer autonomous runs introduced burnout risk plus new bottlenecks in review and decision-making.
Connected Context · Feed7 Judgment

This adds longitudinal, team-level evidence that coding-agent gains depend more on engineering the environment and workflow than on choosing a tool. The spread from under 3x to a 4.5x median among stronger teams narrows broad productivity claims: adoption can begin with a slowdown, and greater autonomy shifts pressure toward specifications, tests, review, decisions, coordination, and sustainable parallel work.

Building a C compiler with a team of parallel ClaudesThe compiler project reinforces the claim that test quality and context hygiene, rather than model capability alone, are central to successful parallel-agent work.Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI LabBoth Amazon accounts frame adoption as maintained team infrastructure, with lean context, automated feedback loops, and repository environment quality shaping results.How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage) — Eyal Blum, FigmaFigma’s emphasis on verification, plans, and review controls matches Amazon’s finding that deployment gains require workflow engineering and can create reviewer pressure.How to build an AI-Native Health Company — Dan Feng, Maven ClinicMaven independently reports the same bottleneck shift toward decisions and verification as implementation accelerates, while both sources limit their figures to local operating contexts.
Context Map
agentcoding#harness-engineering#coding-agents#multi-agent
Uncertainty
The figures come from internal Amazon pilots, including unusually structured experiments, so they may not transfer directly. Teams initially slowed down, and longer autonomous runs introduced burnout risk plus new bottlenecks in review and decision-making.