Amazon will present its framework for engineering trustworthy AI agents at VB Transform 2026
Our take
The rise of AI agents capable of autonomously handling business tasks presents a compelling opportunity, but as the VentureBeat article highlights, a significant hurdle remains: trust. IT leaders are understandably hesitant to grant these agents access to critical enterprise systems, a concern underscored by recent surveys revealing deep anxieties around unauthorized access and prompt manipulation. The prevailing approach of relying solely on model guardrails clearly isn’t enough, with a mere 4% of technology leaders expressing comfort in their efficacy. This hesitation isn’t simply about fear of the unknown; it stems from the inadequacy of current reliability metrics. Traditional EVAL scores offer a limited, static view of performance, failing to capture the crucial aspects of predictability across varied prompts, environments, and input types – a point echoed by Amazon’s Bryan Silverthorn. This resonates with the challenges explored in [How Shopify built an AI stack that doesn't care which models survive], illustrating the need for more robust and adaptable architectures when deploying AI solutions. The simple solution of relying on a single model is increasingly outdated.
Amazon’s proposed framework, centered on consistency, robustness, predictability, and safety, offers a promising direction. Their emphasis on decoupled systems, such as sandboxed environments with human review processes, represents a pragmatic and verifiable approach to building trust, particularly in high-stakes domains like finance. This contrasts with the often-hyped notion of simply “training” AI to be safe, recognizing instead the value of layered security and human oversight. The increased focus on agentic operations and evaluations, as showcased by Waymo’s work in building safe AI for the physical world, which can be explored further in [Intelligence at scale: How Waymo builds safe, efficient AI for the physical world], demonstrates a growing understanding that reliability isn’t an inherent property of a model, but rather a deliberate design outcome. This shift away from chasing “revolutionary” AI toward building incrementally reliable systems aligns with the practical needs of businesses seeking to leverage AI’s potential without undue risk. Even the seemingly simple solution of limiting screen time, as revealed in [If you want to cut your screen time, just get a Brick], demonstrates the power of fundamentally rethinking existing paradigms to achieve practical results.
The significance of Amazon’s approach extends beyond their own internal applications. By openly sharing their framework at VB Transform 2026, they’re contributing to a broader industry conversation around trustworthy AI development. Moving beyond single-agent wrappers to multi-tool architectures that can self-correct during execution represents a significant evolution in agentic AI. This signals a departure from the often-overblown claims of “game-changing” AI and a move towards a more grounded and iterative development process. Companies are realizing that true AI adoption won't be driven by flashy demos, but by demonstrating consistent, predictable, and safe performance—a reliability that builds confidence and ultimately unlocks real business value. The focus on verifiable interactions, even in complex scenarios, is crucial for fostering the widespread acceptance and integration of AI agents into enterprise workflows.
Looking ahead, the ability to quantify and validate AI reliability will be paramount. While Amazon’s framework provides a valuable starting point, developing standardized metrics and auditing processes for agentic AI will be essential for driving further progress. The challenge lies not only in building reliable agents but also in establishing a shared understanding of what "reliable" means in different contexts. As AI agents become increasingly integrated into our business processes, the question becomes not *if* they will make mistakes, but *how* we can design systems that minimize those mistakes and ensure that human oversight remains a critical component of the equation. How will organizations evolve their governance models to accommodate autonomous AI agents operating within complex and sensitive environments, and what new skills will be required to effectively manage and audit these systems?
AI agents are increasingly proficient at executing business tasks autonomously, but IT leaders are cautious about granting permissions to access enterprise systems.
Part of the challenge lies in how AI reliability is measured. Industry standards often rely on EVAL scores, which provide a static snapshot of performance rather than a measure of overall reliability. These metrics can fail to capture predictability across prompts, environments, and input types, said Bryan Silverthorn, director of the AGI Autonomy research lab at Amazon.
Amazon’s AGI autonomy research lab is moving beyond raw performance benchmarks, focusing instead on a structured framework centered on consistency, robustness, predictability, and safety, Silverthorn told VentureBeat during an interview ahead of his session at VB Transform 2026.
Rather than assuming that models can be harnessed into safety, Amazon’s approach emphasizes decoupled systems, such as sandboxed environments where agents propose changes that are reviewed by humans before implementation.
This strategy aims to bridge the trust gap by prioritizing verifiable interactions, even in highly sensitive domains like finance, where the potential damage an agent can cause is significant.
In VentureBeat’s Q2 Pulse Research survey of over 100 senior technology leaders and buyers, just 4% said they are comfortable relying on model guardrails alone. When asked what worries them most about model guardrails, 40% said unauthorized access to tools or data and 27% cited prompt manipulation or injection.
At VB Transform, Silverthorn will share details of Amazon’s approach to trustworthy agentic AI and how companies can move from single-agent wrappers to multi-tool architectures that can self-correct mid-execution during his session titled Closing the capability-reliability gap: Inside Amazon’s framework for engineering trustworthy agents.
Another agentic ops and evals-focused session at VentureBeat’s flagship conference, happening July 14 and 15 in Menlo Park, is Intelligence at scale: How Waymo builds safe, efficient AI for the physical world with speaker Manasi Joshi, director of systems intelligence and machine learning at Waymo.
Interested in attending VB Transform 2026? A select number of complimentary passes are also available to senior technology leaders. Contact us to get yours. You can also purchase tickets here.
Read on the original site
Open the publisher's page for the full experience