Amazon AGI director says AI agent reliability, not capability, is blocking enterprise deployment at VB Transform 2026

Summary

Amazon's AGI director, Bryan Silverthorn, believes that 85% of enterprises pilot AI agents, but only 5% trust them enough to ship to production.
This gap is due to AI agent reliability issues, not their capabilities.
Silverthorn suggests breaking down reliability into four distinct dimensions: consistency, robustness, predictability, and safety.
The framework matters because agents often pass internal evaluations but fail in real-world deployments.
For example, a customer deployed an AI agent for software QA, which worked flawlessly for two months but then began reading wrong numbers due to a software change.
This highlights the need for measurement and evaluation beyond just models.
Silverthorn also emphasizes the importance of understanding the dimensions of variability and matching measurement rigor to the stakes of the application.
His team at Amazon uses a framework that calls AI agents "interns," highlighting the importance of management skills in managing these powerful but occasionally clueless agents.

Why It Matters

Enterprise leaders should prioritize agent reliability over capability when deploying AI agents at scale.
This means understanding the dimensions of variability and matching measurement rigor to the stakes of the application.
Moreover, it's essential to have a clear understanding of the potential risks and consequences of AI agent deployment and to have a plan in place to mitigate these risks.
The lack of enterprise trust in AI agents is a significant issue, as it's hindering the adoption of AI technology.
By prioritizing reliability, enterprise leaders can build trust and ensure that AI agents are used safely and effectively.
Additionally, this emphasis on reliability will drive innovation and improvement in AI technology, making it more suitable for widespread adoption.

GenAI EXPLAINED

An AI agent is a type of software that can perform tasks on its own, often using machine learning algorithms. When we talk about AI agent reliability, we're referring to the consistency and accuracy of these agents' performance. This is different from their capability, which refers to their ability to perform certain tasks.

The concept of breaking down reliability into four distinct dimensions – consistency, robustness, predictability, and safety – is crucial in evaluating AI agents. Consistency refers to the agent's ability to perform tasks consistently over time, while robustness refers to its ability to handle unexpected inputs or situations. Predictability refers to the agent's ability to behave in a predictable manner, and safety refers to its ability to avoid causing harm or damage.

In the context of AI agent deployment, understanding these dimensions is essential in ensuring that the agent is reliable and trustworthy. This involves setting clear expectations, measuring performance, and having a plan in place to mitigate risks.