Six Engineers Instead of 30: How Amazon Rebuilt Bedrock's Inference Engine in 76 Days Using Agentic AI

Summary

The internal Amazon Web Services (AWS) team responsible for rebuilding the Bedrock inference engine successfully delivered the project with just six engineers in 76 days. The project was originally estimated to require 30 developers and take between 12 and 18 months. This massive gain in productivity was achieved by heavily leveraging agentic AI workflows and restructuring the division into highly agile “two-pizza” teams.

What happened?

Successful Rebuilding: A small team of AWS engineers completely rearchitected the core inference engine for Amazon Bedrock in just 76 days.
Dramatic Resource Savings: The project bypassed traditional estimates of 30 developers working for 12 to 18 months, utilizing only six engineers.
Shift in Prototyping Tradition: AWS teams have begun building working demos first rather than starting with Amazon’s sacred “PRFAQ” documents, as AI coding tools make writing code faster than writing long specifications.
Other Examples of Agentic Efficiency:
- Amazon Quick: A desktop application that aggregates searches across mail, Slack, and calendars was built by six engineers in six weeks to beta, reaching 10,000 internal users by week ten, and shipping to public release in three months.
- Strands & Kiro: The AWS software development kit “Strands” and the coding tool “Kiro” were built by small teams, with the Kiro team using Kiro itself to speed up development.

Why it matters

This case study demonstrates a major shift in enterprise software development. By integrating autonomous AI agents, companies can eliminate the communication overhead of large developer teams and return to the “two-pizza team” philosophy with the output of a much larger division. The bottleneck in software development is shifting from writing code to defining exact specifications, designing system architectures, and establishing robust testing environments.

Evidence

GeekWire Investigation: A detailed case study by Todd Bishop explores the organizational shift at Amazon.
AWS Productivity Metrics: In a recent blog post, Swami Sivasubramanian (AWS VP of Agentic AI) noted a median 4.5x productivity increase for teams adapting their processes to AI, with top performers exceeding 10x gains.
Practical Rebuild Example: Sivasubramanian attempted to rebuild a 20-year-old replication engine (used in S3/DynamoDB) using Kiro. After struggling for four nights, he realized the agent lacked testing tools. Once the spec and test suite were set up, the engine was built in two hours.

Analysis

Amazon’s experience shows that simply introducing AI tools to existing rigid processes yields minimal results. Real gains are unlocked only when the organization restructures around the technology—creating smaller teams, promoting cross-functional roles (where product managers write code and engineers make product decisions), and shifting the developer’s role from coding to supervising and testing. Furthermore, continuous testing must be integrated directly into the coding loop so that agents can validate their own outputs.

Practical Takeaways

Reduce Team Sizes: Leverage the productivity gains of AI agents to break down large teams into smaller, highly autonomous units.
Prioritize Test-Driven Specifications: Provide agents with robust testing tools and clear specs beforehand. Without them, agents can get stuck in loops trying to fix errors blindly.
Track AI Token Costs: Monitor and manage token expenses as an operational cost. Well-defined upfront specs minimize wasted token consumption.
Start with the Demo: For low-risk projects, build a functional prototype first to test assumptions rather than spending weeks writing static planning documents.

Open Questions

How will the long-term code quality and maintainability of systems generated primarily by AI agents compare to human-written code?
What new security vulnerabilities might arise from allowing AI agents autonomous test execution and deployment access?
How will organizations measure the value of software development when the marginal cost of coding drops to near zero?