Amazon has unveiled its innovative AI model, Nova Act, designed to create smarter agents capable of performing tasks directly within web browsers. While large language models have popularized the concept of “agents” that mainly respond to queries or retrieve information, Amazon envisions a more advanced role for these agents. The company aims to develop entities that can execute complex, multi-step tasks across various digital and physical environments.
According to Amazon, the ultimate goal for these agents is to handle a wide range of sophisticated tasks, from organizing weddings to managing intricate IT operations, thereby enhancing overall business productivity. However, many existing solutions typically require constant human oversight and extensive API integration, which is not always practical. Nova Act seeks to address these challenges.
In conjunction with the Nova Act model, Amazon is also launching a research preview for the Amazon Nova Act SDK. This toolkit allows developers to create agents capable of automating web tasks such as sending out-of-office notifications, scheduling calendar events, or enabling automatic email replies. The SDK breaks complex workflows into manageable “atomic commands,” enabling fine-tuned instructions to refine task performance.
One of Nova Act’s significant advantages is its remarkable performance on benchmarks. The model has achieved impressive reliability, scoring over 90% in internal evaluations on tasks that often challenge its competitors. For example, it recorded a near-perfect score on the ScreenSpot Web Text benchmark, showcasing its capability in text-based interactions.
Moreover, Nova Act demonstrates adaptability, enabling it to transfer its interface understanding to new environments with minimal training. This versatility allows the model to function effectively in various applications, including enhancing Amazon’s Alexa+ capabilities for self-directed web navigation. Moving forward, Amazon aims to foster the development of intelligent agents trained through reinforcement learning, focused on real-world scenarios that extend beyond simple instructions.
This approach indicates the company’s commitment to transforming the AI agent landscape, paving the way for more complex and practical applications in the future.