Amazon unveils Nova Act, an AI agent that can control a web browser
Amazon on Monday unveiled Nova Act, a general-purpose AI agent that can take control of a web browser and independently perform some simple actions. Alongside the new agentic AI model, Amazon is releasing the Nova Act SDK, a toolkit that allows developers to build agent prototypes with Nova Act.
Nova Act, developed by Amazon’s recently opened San Francisco-based AGI lab, will also power key features of the company’s upcoming Alexa+ upgrade, a generative AI-enhanced version of Amazon’s popular voice assistant. The version of Nova Act available starting today is a little less polished, however. Amazon is calling it a research preview.
Developers can access the Nova Act toolkit on a new website, nova.amazon.com, which also serves as a showcase for Amazon’s various Nova foundation models.
Nova Act is Amazon’s attempt to take on OpenAI’s Operator and Anthropic’s Computer Use with general-purpose AI agent technology of its own. Several leading tech companies believe AI agents that can navigate the web for users will make today’s AI chatbots significantly more useful.
Amazon may not be the first to develop this sort of agentic technology, but via Alexa+, it may have the widest reach.
Amazon says developers building with the Nova Act SDK should be able to automate basic actions on behalf of users, such as ordering salads from Sweetgreen or making dinner reservations. With the Nova Act toolkit, developers can pull together tools that allow an AI agent to navigate web pages, fill out forms, or pick dates on a calendar.
Amazon claims that Nova Act outperforms agents from OpenAI and Anthropic on several of the company’s internal tests. For example, on ScreenSpot Web Text, which measures how an AI agent interacts with text on a screen, Nova Act scored 94%, outperforming OpenAI’s CUA (which scored 88%) and Anthropic’s Claude 3.7 Sonnet (90%).
However, Amazon didn’t benchmark Nova Act using more common agent evaluations, such as WebVoyager.
Nova Act is the first public product to emerge from Amazon’s aforementioned AGI lab, an initiative co-led by former OpenAI researchers David Luan and Pieter Abbeel. Both previously founded startups of their own — Luan started Adept, while Abbeel cofounded Covariant — before Amazon hired them away last year to spearhead its AI agent efforts.
While it may seem strange for an AGI lab to be building AI agents that can order SweetGreen, Luan told TechCrunch that he sees agents as a key step toward creating superintelligent AI systems. Luan defines AGI as “an AI system that can help you do anything a human does on a computer.”
Luan says his team designed the Nova Act SDK to reliably automate short, simple tasks, and give developers tools to precisely define when they want a human to intervene in an agentic workflow. He hopes it will allow developers to create more reliable agentic applications, albeit not necessarily fully autonomous ones.
Amazon is releasing its first generalist AI agent in a crowded space, but it’s a crucial technology that the company has a lot riding on. Early tests of Nova Act could provide a glimpse into some of the capabilities of the long-delayed Alexa+, a make-or-break moment for Amazon’s AI efforts.
A major problem with early AI agents from OpenAI, Google, and Anthropic is their reliability across different domains. In TechCrunch’s tests, the systems are slow, struggle to operate independently for very long, and are prone to mistakes a human wouldn’t make. It won’t be long until we see whether Amazon has cracked the code — or whether its agents suffer from the same flaws plaguing competitors.
Maxwell Zeff is a senior reporter at TechCrunch specializing in AI and emerging technologies. Previously with Gizmodo, Bloomberg, and MSNBC, Zeff has covered the rise of AI and the Silicon Valley Bank crisis. He is based in San Francisco. When not reporting, he can be found hiking, biking, and exploring the Bay Area’s food scene.
