Amazon Launches Nova Act: AI Agent Masters Web Browser Control

Amazon dropped some big news on Monday with the launch of Nova Act, a general-purpose AI agent that can take the reins of a web browser and carry out simple tasks on its own. Alongside this new agentic AI model, Amazon is rolling out the Nova Act SDK, a toolkit that lets developers whip up agent prototypes using Nova Act.

Developed by Amazon's newly opened AGI lab in San Francisco, Nova Act is set to juice up the upcoming Alexa+ upgrade, a souped-up version of Amazon's beloved voice assistant with generative AI smarts. The version of Nova Act hitting the scene today is a bit rough around the edges, though. Amazon's calling it a research preview.

Developers can get their hands on the Nova Act toolkit over at nova.amazon.com, which also doubles as a showcase for Amazon's various Nova foundation models.

With Nova Act, Amazon's throwing its hat into the ring against OpenAI's Operator and Anthropic's Computer Use, aiming to bring its own general-purpose AI agent tech to the table. Big tech players reckon AI agents that can surf the web for users will make today's AI chatbots way more handy.

Amazon might not be the first to jump on this agentic tech bandwagon, but with Alexa+, it could reach the widest audience yet.

Amazon says that with the Nova Act SDK, developers should be able to automate basic stuff for users, like ordering salads from Sweetgreen or snagging dinner reservations. The Nova Act toolkit lets developers cobble together tools that let an AI agent navigate web pages, fill out forms, or pick dates on a calendar.

Amazon claims Nova Act outshines agents from OpenAI and Anthropic in several of its internal tests. For instance, on the ScreenSpot Web Text test, which checks how an AI agent deals with text on a screen, Nova Act scored a solid 94%, beating out OpenAI's CUA (88%) and Anthropic's Claude 3.7 Sonnet (90%).

However, Amazon didn't put Nova Act through the wringer with more common agent evaluations like WebVoyager.

Nova Act marks the first public product from Amazon's AGI lab, a project co-led by former OpenAI researchers David Luan and Pieter Abbeel. Both had their own startups before—Luan with Adept and Abbeel with Covariant—before Amazon scooped them up last year to lead its AI agent charge.

It might seem odd for an AGI lab to be fiddling with AI agents that order salads, but Luan told TechCrunch that he sees agents as a stepping stone to building super-smart AI systems. He defines AGI as "an AI system that can help you do anything a human does on a computer."

Luan says his team built the Nova Act SDK to reliably automate quick, easy tasks and give developers the tools to decide when a human should step in during an agentic workflow. He's hoping it'll help developers create more dependable agentic apps, even if they're not fully autonomous.

Amazon's entering the generalist AI agent game in a packed field, but it's a key tech that the company's betting big on. Early tests of Nova Act might give us a sneak peek at what the long-awaited Alexa+ can do, which could be a make-or-break moment for Amazon's AI ambitions.

One big issue with early AI agents from OpenAI, Google, and Anthropic is their reliability across different areas. In TechCrunch's tests, these systems are slow, struggle to work independently for long, and often make mistakes a human wouldn't. It won't be long before we find out if Amazon's cracked the code—or if its agents have the same issues as the competition.