Amazon has taken a strategic step forward by introducing the Nova Act SDK and launching nova.amazon.com, providing public access to its Nova foundation models. This signals a shift in the company’s AI strategy—from operating as a cloud infrastructure partner to directly equipping developers with tools to build AI agents capable of acting within web browsers.
Nova Act SDK for Web-Based AI Agents
Nova Act is a software development kit (SDK) designed to help developers create agents that can perform human-like tasks within a browser environment. These agents can click buttons, fill out forms, scroll pages, and interact with complex site elements—all through visual understanding and dynamic interaction. Rather than just offering a finished agent, Amazon is giving developers a flexible base to build their own tools.
Google, in comparison, has developed the Chain-of-Agents framework, designed to coordinate multiple AI agents through shared memory and modular communication. Amazon’s Nova Act appear to take a more open-ended route, offering deeper control but demanding more hands-on implementation from developers.
In a demo shared by Amazon, Nova Act demonstrates its browser-agent capabilities by interacting directly with Google Maps. The agent reasons through a task—searching for the “Redwood City Caltrain station”—and visually types the query into the search bar, simulating a human-like action.
On the left side of the screen, a code snippet defines a BikeRoute class, suggesting that the agent will use biking time and distance as constraints to filter apartment listings. This scenario illustrates how Nova Act can interpret user goals and autonomously navigate web interfaces to complete multi-step tasks like planning a commute-centric housing search.
Nova Foundation Models Now Available for Public Use
Amazon has also removed previous barriers to its Nova foundation models by opening up nova.amazon.com, which allows anyone to test and interact with Nova Micro, Lite, and Pro. Previously confined to AWS Bedrock, these models now support public-facing prompts and experimentation—without requiring cloud access or enterprise credentials.
Nova supports 200+ languages and handles contexts up to 300,000 tokens, with plans to reach 2 million tokens later this year.
Earlier this month, Amazon extended its Nova models to AWS GovCloud for use in regulated environments such as government, finance, and healthcare. Developers can also access visual generation tools like Nova Canvas and Nova Reel, which create images and videos and include built-in safety checks and attribution frameworks.
These tools allow developers to trace how visual content is generated, addressing growing concerns around misinformation and synthetic media provenance.
Reasoning Model and Alexa+ Integration Are Next
Looking ahead, Amazon is preparing to release a Nova-branded reasoning model expected by mid-2025. This model will reportedly combine fast conversational capabilities with deeper reasoning, bridging the divide between real-time interactions and long-form analysis.
Amazon is clearly positioning itself to compete with more mature reasoning systems like Claude 3.7 Sonnet, OpenAI’s o3-mini, and the just released Google Gemini 2.5 Pro experimental model.
Meanwhile, Nova Act is expected to play a core role in its new Alexa+ voice assistant, which offers AI-powered automation and seamless service coordination.
Global Competition Highlights Diverging Agent Strategies
While Amazon focuses on tooling, other companies are racing to deliver end-user-facing agents. China’s Zhipu AI just launched AutoGLM, a free agent powered by its lightweight GLM-Z1-Air model.
Designed for constrained environments, AutoGLM runs in-browser or via mobile apps and has been benchmarked—by the company—above GPT-4o and Claude 3.5 Sonnet in Stanford’s AgentBench tests. Zhipu also plans to open-source the agent in April, emphasizing accessibility for developers and global institutions outside the Western AI sphere.
Earlier in the month, Manus AI made headlines for launching a fully autonomous system capable of taking actions without user approval. Built by Butterfly Effect (Hong Kong), the agent employs reinforcement learning, LLM chaining, and a multi-signature control layer to execute workflows and hire contractors.
Following limited beta invites that were resold for thousands of dollars, the company introduced official paid tiers priced at $39 and $199/month.
Amazon’s Full-Stack Ambition Grows
Unlike companies focused on frontend agents, Amazon’s strategy is to build every layer of the AI stack—from custom silicon to foundational models to developer-facing tools. The company’s Nova stack is trained on massive clusters powered by its Trainium 2 chips and backed by billions in infrastructure investment. In a recent interview with Time, AWS CEO Matt Garman emphasized that Amazon’s goal is to offer AI services with long-term cost efficiency and scale.
This vertical integration gives Amazon fine-grained control over model optimization and deployment but also raises the bar for developer adoption. Unlike plug-and-play agents like Operator or AutoGLM, Nova Act requires more effort from users to customize, deploy, and maintain agents at scale.
That trade-off may limit traction among general users but appeal to organizations looking to embed AI deeply within internal workflows or proprietary platforms.
Amazon’s rollout of the Nova Act SDK and public model access reflects a shift toward broader transparency and user control. With its developer-first approach, the company isn’t just enabling AI adoption—it’s empowering a generation of builders to decide what those agents will do and how they’ll do it.
