Amazon Nova Act

What it is

Amazon Nova Act is an AI model that Amazon introduced on 31 March 2025, trained to perform actions inside a web browser rather than only describe them in conversation.^[1] Released alongside it was a research preview of the Nova Act SDK, distributed through nova.amazon.com, which lets developers try building agents that click, type and navigate on a user's behalf.^[1][2]

How it works

The SDK is built around small steps. Rather than asking a model to complete an entire workflow in one shot, developers break the workflow into atomic commands — search for a product, complete a checkout, answer a question about what is currently on the screen — and let the model handle each one.^[1] Because those commands are ordinary Python calls, developers can interleave their own code with browser actions: read a value out of a page, branch on it in Python, then send the agent back into the browser.^[1]

That split also shapes how a build is debugged: each step is a call that can be inspected and retried on its own, so a failure points at a particular action instead of an entire broken automation script. Code handles the control flow that is known in advance; the model handles the page that is not.

The SDK reaches beyond the browser as well. It can call external tools through APIs, through remote MCP, or through agentic frameworks such as Strands Agents, so a browser agent can be one component of a larger system.^[3] On the infrastructure side, the AWS product page describes the service as powered by a custom Nova 2 Lite model and integrated with Amazon Bedrock AgentCore, CloudWatch and IAM — the last of which matters for permissions and auditing in enterprise deployments.^[2] The SDK repository is published under the Apache-2.0 license and carries a warning that versions older than 3.0 are no longer supported.^[3]

Benchmarks

Amazon's case for training a dedicated model is that general-purpose models are not steady enough at this kind of work. The company said state-of-the-art models score between 30% and 60% on benchmarks that measure high-level tasks in web browsers.^[1]

On the narrower job of grounding — finding and identifying an element on a page — the numbers Amazon reported are competitive. On ScreenSpot Web Text, Amazon measured 0.939 for Nova Act, 0.900 for Claude 3.7 Sonnet and 0.883 for OpenAI CUA.^[1] On ScreenSpot Web Icon, Nova Act scored 0.879, and on GroundUI Web it scored 0.805 against 0.825 for Claude 3.7 Sonnet.^[1]

These benchmarks measure whether the model can see and locate things on a page; they are not the same as whether a whole task can be completed, so they are better read as one piece of the picture than as a final verdict.

Where it sits

Nova Act is not only a research artifact. Amazon says the model is already used inside Alexa+, where it navigates the internet when integrated services cannot supply the necessary APIs — the case where a task is possible in a browser but is not exposed as a programmatic interface.^[1]

That placement says something about the design intent: browser agents serve as a fallback layer for the long tail of websites and services that are unlikely to offer an API for every need. Shipping Nova Act as a model plus an Apache-2.0 SDK, wired into Bedrock AgentCore, CloudWatch and IAM, puts it inside Amazon's wider agent tooling rather than presenting it as a standalone finished product.^[2][3]