Amazon Launches Nova Act, An AI Agent for B2B Environments

Amazon AGI SF Lab is introducing Amazon Nova Act, an AI model that claims to outperform competitors such as OpenAI and Anthropic in performance testing for autonomous agents. The technology integrates physical robotics learning and focuses on actionable decisions.

"The basic atomic unit of computation in the future will be a giant AI agent call," David Luan, Head of AGI SF Lab, Amazon, tells Wired.

AI agents, advanced tools capable of performing multiple actions focused on either developing their own or third-party ecosystems, are gaining ground among enterprises. These models, which primarily aim to reduce repetitive administrative burdens such as expense reporting and IT troubleshooting, are allowing organizations to focus on higher-value activities, with companies such as Nvidia, Google, and Microsoft being some of the most advanced in this field.

"At Microsoft, employee success in self-service IT support improved by 36%, while revenue per salesperson increased by 9.4%," says Charles Lamanna, Corporate VP for Business and Industry, Microsoft. "Like any new technology, there will be a learning curve. But if AI agents meet expectations, they could redefine how businesses operate."

However, existing technologies face limitations in reliability and adaptability, reports Wired. Amazon aims to differentiate itself from other developers by focusing on robotic integration and purely business use cases.

Amazon Nova Act Details

The Amazon Nova Act development has been carried out in collaboration with the team of Pieter Abbeel, a pioneer in AI applied to robotics and a professor at the University of California at Berkeley. This synergy, says Amazon Labs, makes it possible to transfer key knowledge from physical systems, such as the robots used in Amazon's distribution centers, to software agents.

The model operates under the premise that if a robot can learn to navigate a warehouse while avoiding unpredictable obstacles, a digital agent can apply similar principles to make decisions in complex virtual environments, such as managing orders or interacting with dynamic web interfaces. This approach not only improves the model's robustness, but also accelerates its ability to operate in real-world scenarios where uncertainty is the norm.

While competitors such as OpenAI and Google focus their efforts on conversational agents, Amazon is betting on actionable automation. A shared example focuses on e-commerce, with Nova Act being able to manage shopping carts by analyzing a user's history, but goes further by integrating safeguards thanks to its specialized software development kit (SDK).

The model has already demonstrated its efficiency in technical support in tasks such as scheduling repairs through assistants like Alexa, reducing the operational burden on human teams, reports Amazon. Moreover, its potential in logistics, optimizing inventories in real time using data from warehouse robots, underscores its design to solve concrete problems, not just generate answers.

In logistics, integration with Amazon's robotic systems enables inventory adjustment based on predictive demand. Each use case is designed to scale without requiring constant monitoring, a key requirement for enterprise adoption.

Its standardized test results reflect the technical robustness of Amazon Nova Act. In the GroundUI Web test, the model achieved 94.3% accuracy, beating Anthropic's Claude 3.7 Sonnet by more than five percentage points. In ScreenSpot, which evaluates the ability to navigate complex interfaces, it scored 88.7%, three points above OpenAI's CUA Model, reports Amazon.

According to both tests, these numbers were achieved because the model is based on an optimized version of Amazon Nova, specifically trained to decide when to act and how to act, minimizing reliance on manual rules that limit the flexibility of other agents. This approach reduces errors in open tasks, a critical pain point for companies that rely on reliable automation.

Amazon is also offering the tools to integrate it. Its SDK includes plug-and-play capabilities that simplify the creation of agents capable of interacting with interfaces designed for humans, such as web forms or dynamic menus. This lowers barriers to entry for companies that lack specialized AI teams.

According to Luan, the next 18 months will focus on improving fault tolerance, especially for sensitive sectors such as finance or healthcare, where minor errors have serious consequences. 

Amazon is channeling its record US$12 billion investment in AI into hands-on automation, a niche with growing demand in sectors such as retail and manufacturing. Analysts say that this strategy could give it an edge in markets where operational efficiency is a priority. The challenge will be to demonstrate that its agents can maintain high standards of reliability at scale.