Amazon Web Services rebuilt Bedrock's inference engine with just six engineers in 76 days.
The Information reported that the sprint replaced an original plan calling for 30 to 40 engineers working over 12 to 18 months.
How the smaller team pulled it off
The compressed team built a new runtime called Mantle to power the rebuilt engine.
They relied heavily on an internal agentic coding tool named Kiro, which allowed the small group to move at a pace closer to that of a much larger team.
What the rebuild was trying to fix
The work targeted persistent latency and throughput bottlenecks in Bedrock's inference path, the process by which the platform actually runs AI models for customers.
Fixing that bottleneck has become a priority as AWS has broadened Bedrock's model catalogue and agent features, and begun hosting OpenAI's models alongside its own.
That expansion has narrowed a point of differentiation that had previously favoured Microsoft's rival Azure platform.
How AWS is framing the win
Internally, AWS has framed the effort as both a productivity and an operational win.
The company says the rebuilt engine enables higher-volume, lower-latency managed agents and production inference for enterprise customers.
A template for what comes next
The Information reported that the team's approach will now be used as a template for further upgrades to Bedrock.
That signals a broader push by AWS to compete on agent infrastructure, governance and platform breadth, rather than relying on exclusive access to any single AI model.
