Amazon Web Services has completed a major overhaul of the inference engine for its Bedrock AI service. The new architecture, known as Project Mantle, was developed to address performance and capacity constraints caused by a surge in AI workloads.
The project was completed in 76 days by a team of six engineers. Initial internal projections estimated the task would require 30 to 40 engineers working over a period of 12 to 18 months.
The team utilized Kiro, an internal agentic AI coding tool, to generate specifications, code, tests, and deployments. AWS leadership reported that using such tools can increase engineer productivity by 10 to 20 times.
Despite the reliance on AI-generated code, the company maintains a strict policy requiring human review for every line of code before it is deployed. CEO Andy Jassy highlighted the project in his 2025 shareholder letter.
The Mantle architecture is now live and operates through a dedicated service endpoint, separate from the legacy runtime. This transition follows a period of rapid growth, where Bedrock usage nearly doubled in March 2026.
