Summary
Clare Liguori, senior principal engineer at AWS, presents a data-backed case for what Amazon internally calls “frontier development” — a new phase of AI-assisted engineering defined by hands-off agent operation, minimal human intervention, and running multiple agents in parallel. Based on pilots across Amazon, she reports a median 4.5x productivity improvement with some teams reaching 10x, a step-function change beyond the 10–20% gains seen from earlier AI coding assistance.
Three case studies provide concrete grounding. The Bedrock Mantle team rebuilt Amazon’s model hosting inference data plane — originally scoped at 30 engineers over 18 months — with six engineers in 76 days using Kiro, AWS’s agentic coding assistant. A Prime Video experimental sprint compressed a 90-week delivery estimate to 24 weeks in just 10 days. Amazon Stores then ran a broader structured pilot across 50 normal teams to validate these results beyond controlled sprints.
Liguori identifies the key behavioral shifts required for frontier development to work: writing detailed steering files that externalize institutional knowledge agents would otherwise lack; “slowing down to speed up” by investing in agent context before expecting productivity gains (improving error messages, building MCP servers, restructuring codebases, and in some cases switching to TypeScript or Rust for better compiler feedback); and treating agents as parallel workers to feed with well-scoped tasks rather than assistants to babysit turn-by-turn. The talk also addresses how steering file hygiene evolves as model capabilities improve, noting that workarounds needed for Sonnet 3.7 are often unnecessary with Opus 4.5 and later.
📺 Source: AI Engineer · Published August 28, 2026
🏷️ Format: Workflow Case Study
