What you need to know about Amazon Nova Act: the new AI agent SDK challenging OpenAI, Microsoft, Salesforce

Sleepy giant woke up!

For a while, it seemed to be playing in the competition in the race to present to Amazon users – especially in the structure of millions of developers Amazon Web Services (AWS) ‘s Cloud infrastructure – first party attracts AI models and tools.

But at the end of 2024 this His internal basic model family made the Amazon Nova debutWith text, image and even video generation capabilities and Last month he saw a new Amazon Alexa voice assistant The anthropick is partially intensified by the claw family.

Then, Monday, e-commerce and cloud giant Artificial General Intelligence Division Amazon Agi have Amazon announced the release of the Nova ActAn experimental developer set to create AI agents that can navigate the Internet can complete autonomy that can be equipped with a special version of Amazon’s Nova Model (LLM). OH, and Standard Developer Kit (SDK), although the SDK is an open source under a Permisive Apache 2.0 license, is designed to work with the Special Nova model of the Amazon model that is not only a third party.

The goal is to allow third party developers to build AI agents, who are able to perform safely within the web browsers.

But how does Amazon’s Nova Accoru, for example, to the construction platforms in which there are other agent construction platforms? Microsoft’s autogenous, Salesforce’s Agent Forceand, of course, Openai recently open source Agents SDK?

A different, thoughtful approach to AI agents

Since the public rise of major language models (LLS), most agent systems have been limited to respond to natural languages ​​or asking for knowledge bases.

NOVA ACT is part of a larger industry of agent systems based on the user-based activity that can perform the actual tasks in the digital environments. Openai’s new answers APIThis is an example of the development of the developers, which allows users to access the autonomous browser navigator.

Amazon Agi requires human control when the current agent systems, prospects and promising and often require human control, especially when working multiple steps or complex work flows.

NOVA ACT is designed to solve these restrictions by presenting these restrictions to reliable workflows, prescribing prescription commands.

Deniz Birliki, a member of the technical staff in Amazon, explained a broader vision in A The video presents the Nova Act: Soon they will have more AI agents than people who perform the tasks on behalf of users.

https://www.youtube.com/watch?v=jllapxwalu

The Amazon Autonomy Board and the VP head of the AGI SF Laboratory David Luan, in the last video call interview with VentureBeat, we further increased the mission.

Luan, before Adept EU co-founder and CEOJoined Amazon in 2024 A part of AIGTUI-Rent. Luan has long been a supporter of AI agents. “It’s easy, we were the first company that really started working on AI agents. At this point, everyone knows how important it is.

Nova ACT is safe

NOVA ACT SDK provides a frame for the construction of web-based automation agents using natural language offers divided into clean, managed steps on the Internet.

Unlike typical LLM-energies, which results in a frequent behavioral act of approximately an invalid behavior-nova act, it is designed to re-execute even smaller, checkable tasks.

Some of the main features of the Nova Act are:

• Slim title task breakdown: Developers can reduce the complex digital work flows that lead the agent to the agent, each agent, to fulfill its special UI interaction.
• Direct browser manipulation via playwright: Nova combined with ACT Playwrightdeveloped by an open source browser automation framework Microsoft. Playwright allows developers to manage web-browsers program, filling elements, filling forms or navigation pages, only confidence in AI forecasts. It is especially useful for managing sensitive tasks such as integration, passwords or credit card information. For example, instead of sending sensitive information to the model, developers can use the password to a password area and then to enter the password to securely enter the password without “seeing”. This approach helps strengthen security and privacy while automating the web interaction.
• Python Integration: SDK allows developers to combine Python code with standard python tools such as Python Tools, Key Pots, claims or ropes for parallel execution, including NOVA Acton commands.
• Structured data extraction: The SDK supports the removal of structured data through Pydantic Schemas, which allows agents to convert screen content to structured formats.
• Dealing and Planning: Developers can operate more than one Nova ACT instance at the same time and plan automated work flows without the need for continuous human control.

Luan stressed that the Nova Act was a tool for developers more than a common conversation. “NOVA ACT is set for developers. This is not a conversation that you are talking to this fun. It is designed to allow developers to start building useful products.”

For example, one of the workflows shown in Amazon’s documents, how to use the rental lists of Nova ACT by calculating the scorching distance and ride a bicycle.

Another exhibited example uses a Nova’s law to order a special salad from Sweetgreen, which can automate the digital tasks that can automate the digital tasks in a table every Tuesday, delicate and customizable, confident and customizable.

Attention to evaluation performance and reliability

A central message in Amazon’s announcement is the main obstacle to adoption of reliability, not just intelligence.

Advanced models, AI agents, AI AI agents, according to Amazon, are quite fragile with EU agents that have a very step-by-step tasks that are typically from 30% to 60%.

At the same time, NOVA Emphasizes a building block that challenges other models, other models, such as ACT, other models, openings, openings, interior voters or pop-ups.

Luan stressed why this reliability is important. “How do we really make our attention really true how reliable agents?

Amazon AGI, NOVA is against competitive models, including an ancient, Anthropic Claude 3.7 Sonet and Openai’s Cua model. Instructions in the NOVA Act, the screenshot of the screenshot that tests the screenshots, Nova ACT 0.939, Claude 3.7 Sonnet (0.900) and Openai Cua (0.883) collected a score.

At the screenshot of the visual UI elements, on the screenshot of the web icon, Nova ACT, other models hit 0.879 again and again.

However, General UI has hit a benching on a surface website, which tests the mutual connection, NOVA ACTI, hit 0.805 behind their opponents.

These scores were consistently measured internally by Amazon using the guidelines and evaluation criteria.

Amazon also stressed early results in the ability to generalize the NOVA ACT outside the standard environment.

For example, the team member Rick Liu demonstrated how the statistics, which determined the pigeon-themed web game, without learning an agent, a pigeon-themed-themed website game, and how to fight competitors and progress in the game.

According to Luan, the ability to generalize is the center for long-term vision. “Our goal with the NOVA Act is to use a universal browser. We’re an agent who can do everything you want to do on a computer on a computer,”

Flexible for use in different clouds, but locked in the Amazon’s Nova

The law of Nova is available for global development nova.amazon.comLuan clarified that the system was integrated in the Amazon’s house in Nova Foundation.

They cannot connect to external LLMS like creators Openai’s GPT-4O or Anthropic Clod 3.7 SonnetUnlike Openai agents, SDK and less extent, Microsoft’s autogenous and Salesforce’s Agent Force Platforms (platforms that allow several different providers and model families).

“NOVA ACT is a specially designed version of the Nova model,” he said. “This is not only a gross-generic llm. Indeed, it is trained to move on the Internet from your behalf.”

However, the NOVA law is not limited to the AWS environment. Developers can download SDK and manage locals, in clouds or choices. “You don’t have to be in AWS to use it,” Luan said.

Thus, for businesses looking for convenience of maximum main model for agents, Nova ACT is probably not the best choice. However, for a model looking for a model designed specifically designed to walk on the Internet and move on different websites (UIS), you will probably need a look – especially if you are in the Amazon or AWS developer ecosystem.

Security, licensing and price

Nova ACT SDK, Apache License, version 2.0 (January 2004), is broadcast under the open source license. However, this applies only to the SDK program.

The Nova Act remains a property and closed source with itself with itself, weight and training information. Explaining that the model is reliably integrated and the SDK is integrated with a solid and associated with the SDK, it is intentionally intentional.

In the beginning, NOVA ACCORE is offered as a free research review. There is still no declared price for production.

Luan described this stage as an opportunity for developers to build with experience and technology. “Our faith is not yet established yet.

For a longer period, Amazon plans to present production-based shipping and scaling guarantees, but these are not yet available.

What is for the NOVA law?

The release of the NOVA Act, Amazon’s action-oriented AI agents reflect a broader passion for preparing a component component.

Luan concluded the opportunity: “My personal dream is that agents are becoming a building block of the computing block, and the best new beginnings and products are built on the development of our team.”

Nova ACT is already available for SDK, experience and prototiping Amazon’s website and Entrusted.