The enterprise AI industry faces a significant challenge. According to data from Cisco, while 85% of enterprises are piloting AI agents, only 5% have integrated them into production environments. At **VB Transform 2026**, Brian Silverthorne, Director of AGI Autonomy at Amazon, shared insights into this gap and explained why enhancing benchmarks isn’t the solution.
Having joined Amazon through the acquisition of Adept AI, Silverthorne now leads multimodal agent training at the company’s AGI Lab. He highlighted the importance of reliability, which he breaks down into **four dimensions**: **consistency, robustness, predictability, and safety**. This framework is based on studies conducted at Princeton University.
“This unpacks the various factors intertwined in almost every evaluation I have encountered,” he stated.
Why do AI agents succeed in internal assessments yet struggle with real customers?
This framework is crucial as AI agents often undergo periodic internal evaluations before experiencing failures in real-world applications. Silverthorne recounted a case where a customer deployed an agent for software QA that extracted serial numbers from screens. While it performed well for two months, it began misreading numbers sporadically. The culprit? The vision encoder’s performance varied based on the serial number’s screen location, and a subtle software change led to its downfall.
Silverthorne emphasized that this lesson revolves around measurement, not just model enhancement. “The model must improve. We are dedicated to enhancing the model,” he stated. However, the more profound lesson is for teams to identify variability dimensions and align their measurement rigor with application interests. VentureBeat’s prior research supports this, revealing that half of surveyed companies deploy agents that succeed in evaluations but fail real users. Many organizations overlook accuracy by primarily monitoring uptime and conducting superficial checks without thorough diagnostics, leading to minimal guardrails in place. Most default to the model maker’s ratings, relying heavily on either trusting the vendor or taking a gamble.
Understanding Amazon’s “Intern” Framework for Autonomous AI Management
Silverthorne’s most impactful insights focused on cultural practices rather than technical solutions. In Amazon’s AGI Lab, researchers colloquially refer to agents as **“interns.”** This reflects the underlying management philosophy where agents are seen as powerful yet sometimes naive, capable of both impressive achievements and significant failures.
He proposed that managing these systems requires managerial skills rather than purely technological ones. This includes contemplating potential pitfalls, establishing backup strategies, and consciously evaluating acceptable risks. “You can approach the interns and ask, ‘What could you do wrong? How can we mitigate negative outcomes?'” he explained. Amazon’s labs acknowledge the trade-off, prioritizing research speed even if it occasionally leads agents to conduct flawed experiments, such as running continuous experiments based on their own high-level research agendas.
Essential Steps for Business Leaders Before Scaling AI Agent Deployment
Silverthorne candidly expressed the limitations of current technology. He deemed the term “self-improving AI” an “overused phrase.” While Amazon actively employs AI to enhance its models, it remains far from achieving complete autonomous self-improvement. The focus within his lab continues to be on computing, as commercial trucking clients utilize browser automation to integrate warranty claims across disjointed systems. He emphasized that future agents will not only depend on computing but will also collaborate with MCPs, APIs, and other tools to execute end-to-end workflows. Although employing LLMs as evaluators presents potential, it represents just one of several strategies for balancing agent capabilities with acceptable risks.
For organizations trapped in a cycle of experimentation, the way forward requires a **mindset shift.** Instead of questioning if agents can achieve a singular, exceptional task, they should focus on whether these agents can execute the task accurately **1,000 times** in succession.
Ultimately, the organizations capable of transcending the **85% deployment hurdle** won’t just have the smartest AI agents; they’ll have the most effective managers.
Source: venturebeat.com