Amazon has introduced Nova Act, a general-purpose AI agent designed to use a web browser and complete some simple actions on its own. The release also includes the Nova Act SDK, giving developers a way to build early agent prototypes around the model.
What Nova Act Is Built To Do
Nova Act is Amazon’s entry into a fast-moving category of AI systems that do more than generate answers. Instead of only responding in a chat window, an agent like Nova Act can interact with pages, move through web flows, and carry out basic tasks for a user.
Amazon says developers using the Nova Act SDK should be able to automate simple actions such as ordering salads from Sweetgreen or making dinner reservations. The toolkit is designed to let an agent navigate web pages, fill out forms, and pick dates on a calendar.
That makes Nova Act part of a broader shift in AI product design. The core idea is that a useful assistant should not just explain what to do; it should be able to do some of the work inside the browser. Several leading tech companies are working on this same problem because web navigation is central to how people use computers.
A Research Preview With Alexa+ Ambitions
The version of Nova Act available now is not being presented as a finished consumer product. Amazon calls it a research preview, which signals that the company is putting the technology in developers’ hands while it is still being tested and refined.
Developers can access the Nova Act toolkit at nova.amazon.com. That site also serves as a showcase for Amazon’s Nova foundation models.
Nova Act was developed by Amazon’s recently opened San Francisco-based AGI lab. The model is also expected to power key features in the company’s upcoming Alexa+ upgrade, a generative AI-enhanced version of Alexa.
That connection gives Nova Act more strategic weight than a standalone developer experiment. Amazon may not be first to this kind of browser-using agent technology, but Alexa+ could give the company a large path to users if the agent features work well enough.
How Amazon Frames The Competition
Nova Act arrives in a market where OpenAI and Anthropic have already shown their own approaches to computer-using agents. Amazon is positioning Nova Act against OpenAI’s Operator and Anthropic’s Computer Use.
Amazon says Nova Act performs strongly on some of its internal tests. On ScreenSpot Web Text, which measures how an AI agent interacts with text on a screen, Nova Act scored 94%. Amazon said OpenAI’s CUA scored 88%, while Anthropic’s Claude 3.7 Sonnet scored 90%.
Those numbers are notable, but the comparison has limits. Amazon did not benchmark Nova Act using more common agent evaluations, such as WebVoyager. That matters because agent performance can look different depending on the task, the benchmark, and the kind of browser behavior being measured.
For developers, the practical question is not only whether an agent can do well on a screen-text test. It is whether it can handle everyday workflows reliably enough to be useful, predictable, and easy to supervise.
The AGI Lab Behind The Product
Nova Act is the first public product from Amazon’s AGI lab. The lab is co-led by former OpenAI researchers David Luan and Pieter Abbeel.
Both had founded startups before joining Amazon. Luan started Adept, while Abbeel co-founded Covariant. Amazon hired them last year to lead its AI agent work.
Luan connects the browser-agent work to a much larger ambition. He told TechCrunch that agents are an important step toward superintelligent AI systems, and he defines AGI as “an AI system that can help you do anything a human does on a computer.”
That definition helps explain why a system that can order salads or make reservations matters to Amazon’s research agenda. In this view, simple browser tasks are not trivial side projects. They are early tests of whether an AI system can operate software in the same environment people use every day.
Reliability Remains The Central Question
Amazon’s release focuses on short, simple tasks rather than full autonomy. Luan says the Nova Act SDK was designed to automate those narrow workflows reliably and to let developers specify when a human should step in.
That design choice is important because reliability has been a major weakness for early AI agents from OpenAI, Google, and Anthropic. In TechCrunch’s tests, those systems were slow, had difficulty working independently for long periods, and made mistakes that a person would not usually make.
Nova Act’s real test will be whether Amazon can reduce those problems in practical use. Browser agents need to interpret pages, make choices, complete forms, and recover when something unexpected appears. Even simple tasks can become difficult when a website changes or a workflow requires judgment.
For now, Nova Act gives developers a preview of how Amazon wants agentic AI to work: narrow enough to supervise, useful enough to automate common actions, and closely connected to the future of Alexa+. The research preview should reveal whether Amazon has made meaningful progress on the reliability challenge or whether its agents face the same limits as competitors.