en.Wedoany.com Reported - Bryan Silverthorn, Director of Amazon's AGI Autonomy research lab, recently stated that the way to measure the reliability of AI agents is shifting from static performance metrics to a focus on consistency, robustness, predictability, and safety. Silverthorn pointed out that the EVAL scores relied upon by industry standards only provide a static snapshot of performance, making it difficult to fully reflect an AI system's predictability across different scenarios. This poses a major challenge for enterprises when granting system access to AI agents.
Silverthorn spoke in an interview ahead of the upcoming VB Transform 2026 conference, scheduled for July 14-15 in Menlo Park. He revealed that Amazon's strategy does not assume the model itself is safe and reliable, but instead emphasizes a decoupled system architecture. Specific measures include introducing a sandbox environment where AI agents can propose changes within a controlled scope, which are then reviewed by humans before execution. This approach aims to bridge the trust gap in enterprise applications by prioritizing verifiable interactions, particularly in high-risk, sensitive fields such as finance.
According to VentureBeat's Q2 Pulse Research survey, among over 100 senior technology leaders and buyers, only 4% are willing to rely solely on model guardrails to ensure AI safety. Among the top concerns of respondents, 40% worry about unauthorized access to tools or data, while 27% cited risks of prompt manipulation or injection. Silverthorn plans to share details at VB Transform 2026 on how Amazon is building trustworthy Agentic AI, and will introduce the transition from single-agent encapsulation to a multi-tool architecture capable of self-correction during execution.
The conference will also feature other talks focused on Agentic operations and evaluation, such as Manasi Joshi, Director of System Intelligence and Machine Learning at Waymo, who will share the company's experience in building safe and efficient AI in the physical world.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
