Developed at Amazon's newly opened AGI Labs in San Francisco, NovaAct will also provide key functionality for the company's upcoming upgrade to Alexa+, a generative AI enhancement to Amazon's popular voice assistant. However, the version of NovaAct available starting today is slightly inferior. Amazon calls it a research preview.
Developers can access the NovaAct toolkit through the new website nova.amazon.com, which also serves as a showcase for Amazon's various Nova base models.
NovaAct is Amazon's attempt to leverage its own general artificial intelligence agent technology to compete with OpenAI's Operator and Anthropic's ComputerUse. Several leading tech companies believe that AI agents capable of navigating the web for users will make today's AI chatbots more useful.
Amazon may not be the first to develop this kind of agent technology, but with Alexa+, its reach may be the broadest.
Amazon says developers building with the NovaAct SDK should be able to automate basic actions on behalf of users, such as ordering a salad from Sweetgreen or making a dinner reservation. With the NovaAct toolkit, developers can integrate tools that let AI agents browse the web, fill out forms, or select dates on a calendar.
Amazon claims that NovaAct outperformed OpenAI and Anthropic's agents in several tests within the company. For example, in ScreenSpot WebText, which measures how an AI agent interacts with text on the screen, NovaAct scored 94%, better than OpenAI's CUA (scoring 88%) and Anthropic's Claude3.7Sonnet (90%).
However, Amazon did not benchmark NovaAct using more common proxy evaluations such as WebVoyager.
NovaAct is the first public product launched by Amazon's above-mentioned AGI laboratory. The project is co-led by former OpenAI researchers David Luan and Pieter Abbeel. Both have previously founded their own startups — Luan founded Adept, while Abbeel co-founded Covariant — and Amazon hired them last year to lead its AI agent efforts.
While it may seem odd that AGI Labs would develop an AI agent capable of ordering SweetGreen, Luan believes agents are a critical step in creating super-intelligent AI systems. Luan defines AGI as "an AI system that can help you do everything a human does on a computer."
Luan said his team designed NovaActSDK to reliably automate short tasks and provide developers with tools that allow them to define precisely when human intervention is needed in agent workflows. This, he hopes, will allow developers to create more reliable proxy applications, although not necessarily fully autonomous ones.
Amazon has launched its first general artificial intelligence agent in a crowded market, but it's a key technology the company has high hopes for. Early testing of NovaAct offers a glimpse into some of the capabilities of the long-delayed Alexa+, marking a make-or-break moment for Amazon's artificial intelligence efforts.
The main issue with early AI agents from OpenAI, Google, and Anthropic is their reliability in different domains. In TechCrunch's testing, these systems were slow, difficult to run independently for long periods of time, and prone to making mistakes humans wouldn't make. We'll soon see if Amazon has cracked the code -- or if its agents suffer from the same flaws that have plagued its competitors.