AWS: Nearly 90% of Amazon’s Early AI Agent Prototypes Never Shipped
AWS vice president Swami Sivasubramanian told HumanX that almost 90% of Amazon's early AI agent prototypes never reached production. He detailed how Amazon addressed this issue, from Kiro to Bedrock.
Key Takeaways:
- Deployment Challenges: Only 17% of organizations successfully deploy AI agents, and only 7% can measure the return on investment (ROI), according to Sivasubramanian.
- Root Causes of Failure: Five main reasons why AI agent projects stall include incorrect problem definition, inability to measure success, inadequate governance, leadership indecisiveness, and a lack of organizational redesign for AI.
- Amazon's Approach: Amazon allowed numerous prototypes to flourish and compete, measuring their performance. This led to the open-source project Kiro Crew, with close to 39,000 employees contributing within 30 days of its launch.
- Kiro and Bedrock: To streamline security and infrastructure concerns, Amazon consolidated onto Bedrock for AI inference and AgentCore for agent hosting.
- Strands Framework: AWS now offers an open-source agent framework called Strands with a feature that places an agent in a deterministic "box" for governance.
Date: September 24, 2026 Source: HumanX Conference, Amsterdam