Amazon Nova Act: AI Agent Controls Web Browser

Amazon Introduces Nova Act: A New General-Purpose AI Agent

On Monday, Amazon announced the release of Nova Act, a versatile AI agent designed to manage a web browser and execute straightforward tasks autonomously. Accompanying this new model is the Nova Act SDK, a development toolkit enabling creators to prototype agents utilizing Nova Act’s capabilities.

Developed by Amazon’s AGI Lab

Nova Act originates from Amazon’s recently established AGI laboratory located in San Francisco. This agent will also be integral to the functionality of the forthcoming Alexa+ upgrade, an enhanced version of Amazon’s widely-used voice assistant powered by generative AI.

Currently, the available version of Nova Act is presented as a research preview, indicating it is still undergoing refinement.

Accessing the Nova Act Toolkit

Developers can gain access to the Nova Act toolkit through a dedicated website, nova.amazon.com. This platform also showcases Amazon’s diverse Nova foundation models.

Competing in the AI Agent Landscape

Amazon’s introduction of Nova Act represents its entry into the competitive field of general-purpose AI agent technology, directly challenging offerings like OpenAI’s Operator and Anthropic’s Computer Use. A growing consensus among leading technology firms suggests that AI agents capable of web navigation will significantly enhance the utility of current AI chatbots.

While not the first to venture into this technology, Amazon’s integration with Alexa+ potentially provides it with the broadest user reach.

Capabilities and Potential Applications

Amazon states that developers utilizing the Nova Act SDK will be able to automate routine tasks for users. Examples include ordering meals from Sweetgreen or arranging dinner reservations.

The toolkit empowers developers to create tools that allow an AI agent to navigate web pages, complete forms, and select dates from calendars.

Performance Benchmarks

Amazon asserts that Nova Act surpasses agents from OpenAI and Anthropic in several internal evaluations. Specifically, on the ScreenSpot Web Text benchmark – measuring AI agent interaction with on-screen text – Nova Act achieved a score of 94%, exceeding OpenAI’s CUA (88%) and Anthropic’s Claude 3.7 Sonnet (90%).

However, Amazon did not evaluate Nova Act using more prevalent agent benchmarks like WebVoyager.

Origins in Amazon’s AGI Initiative

Nova Act is the inaugural public product stemming from Amazon’s AGI lab, a project jointly led by former OpenAI researchers David Luan and Pieter Abbeel.

Both Luan and Abbeel previously founded their own companies – Adept (Luan) and Covariant (Abbeel) – before joining Amazon last year to lead its AI agent development.

The Role of Agents in Achieving AGI

Luan explained to TechCrunch that he views AI agents as a crucial stepping stone towards the creation of superintelligent AI systems. He defines AGI as “an AI system capable of assisting with any task a human can perform on a computer.”

The Nova Act SDK was designed by his team to reliably automate concise, uncomplicated tasks, while also providing developers with the means to specify when human intervention is required within an agentic workflow.

The goal is to facilitate the development of more dependable agentic applications, even if they are not entirely autonomous.

A Critical Technology for Amazon

Amazon is entering a competitive market with its first generalist AI agent, but this technology is vital to the company’s future. Initial testing of Nova Act may offer insights into the capabilities of the long-awaited Alexa+, a pivotal moment for Amazon’s AI endeavors.

Addressing Reliability Concerns

A significant challenge with early AI agents from OpenAI, Google, and Anthropic is their inconsistent performance across different applications. Tests have revealed that these systems can be slow, struggle with prolonged independent operation, and are susceptible to errors that humans would readily avoid.

It remains to be seen whether Amazon has overcome these limitations, or if its agents will exhibit similar shortcomings to those of its competitors.