Amazon on Monday unveiled Nova Act, a general-purpose AI agent that may take management of an online browser and independently carry out some easy actions. Alongside the brand new agentic AI mannequin, Amazon is releasing the Nova Act SDK, a toolkit that enables builders to construct agent prototypes with Nova Act.
Nova Act, developed by Amazon’s lately opened San Francisco-based AGI lab, may also energy key options of the corporate’s upcoming Alexa+ improve, a generative AI-enhanced model of Amazon’s common voice assistant. The model of Nova Act obtainable beginning at the moment is rather less polished, nonetheless. Amazon is looking it a analysis preview.
Builders can entry the Nova Act toolkit on a brand new web site, nova.amazon.com, which additionally serves as a showcase for Amazon’s varied Nova basis fashions.
Nova Act is Amazon’s try and tackle OpenAI’s Operator and Anthropic’s Laptop Use with general-purpose AI agent expertise of its personal. A number of main tech firms imagine AI brokers that may navigate the net for customers will make at the moment’s AI chatbots considerably extra helpful.
Amazon is probably not the primary to develop this form of agentic expertise, however by way of Alexa+, it might have the widest attain.
Amazon says builders constructing with the Nova Act SDK ought to be capable to automate primary actions on behalf of customers, equivalent to ordering salads from Sweetgreen or making dinner reservations. With the Nova Act toolkit, builders can pull collectively instruments that enable an AI agent to navigate net pages, fill out types, or choose dates on a calendar.
Amazon claims that Nova Act outperforms brokers from OpenAI and Anthropic on a number of of the corporate’s inside exams. For instance, on ScreenSpot Internet Textual content, which measures how an AI agent interacts with textual content on a display screen, Nova Act scored 94%, outperforming OpenAI’s CUA (which scored 88%) and Anthropic’s Claude 3.7 Sonnet (90%).
Nonetheless, Amazon didn’t benchmark Nova Act utilizing extra frequent agent evaluations, equivalent to WebVoyager.
Nova Act is the primary public product to emerge from Amazon’s aforementioned AGI lab, an initiative co-led by former OpenAI researchers David Luan and Pieter Abbeel. Each beforehand based startups of their very own — Luan began Adept, whereas Abbeel cofounded Covariant — earlier than Amazon employed them away final yr to spearhead its AI agent efforts.
Whereas it might appear unusual for an AGI lab to be constructing AI brokers that may order SweetGreen, Luan advised TechCrunch that he sees brokers as a key step towards creating superintelligent AI programs. Luan defines AGI as “an AI system that may provide help to do something a human does on a pc.”
Luan says his crew designed the Nova Act SDK to reliably automate brief, easy duties, and provides builders instruments to exactly outline when they need a human to intervene in an agentic workflow. He hopes it’s going to enable builders to create extra dependable agentic purposes, albeit not essentially totally autonomous ones.
Amazon is releasing its first generalist AI agent in a crowded house, however it’s a vital expertise that the corporate has quite a bit using on. Early exams of Nova Act might present a glimpse into a number of the capabilities of the long-delayed Alexa+, a make-or-break second for Amazon’s AI efforts.
A serious drawback with early AI brokers from OpenAI, Google, and Anthropic is their reliability throughout totally different domains. In TechCrunch’s exams, the programs are sluggish, battle to function independently for very lengthy, and are susceptible to errors a human wouldn’t make. It received’t be lengthy till we see whether or not Amazon has cracked the code — or whether or not its brokers endure from the identical flaws plaguing rivals.