Web browser operation automation AI agent can be built! Amazon Nova Act is now generally available #AWSreInvent
This page has been translated by machine translation. View original
I'm Ashizawa.
I am attending AWS re:Invent 2025 in person.
On December 2 (local time), Amazon Nova Act became generally available (GA).
In this article, I will introduce an overview of the service and briefly share my testing experience.
Overview
Amazon Nova Act is a service for building and managing reliable AI agents that perform actions in web browsers.
It can reliably complete repetitive UI workflows in browsers, execute APIs or tools (such as writing to PDFs), and implement Human-in-the-loop capabilities to escalate to humans when necessary.
This service was announced as a research preview in January 2025, expanded to limited preview in July 2025, and has now been made generally available.
The YouTube video posted during the limited preview makes it easier to understand. Please take a look.
Usage Flow
You can experience UI workflow automation agents created with Nova Act by using the Nova Act Playground available at https://nova.amazon.com/act.
To use the Playground, you need to log in to the Amazon Nova site, and a US Amazon.com account is required as the login user. Note that you cannot log in with a Japanese Amazon.co.jp account.
First, enter the following information:
- Agent name
- URL of the site to be operated
- Description of the agent's actions
Since I didn't have a test environment, I used the pre-defined preset "Hello World" for this test.
- Site URL:
https://nova.amazon.com/act/gym/next-dot/search - Agent action: Find flights from Boston to Wolf on Feb 22nd
It appears that the agent will use a fictional travel booking website to search for flights.
When you click the right arrow (Run Agent), the agent runs and you can see the agent operating the browser.
On the right side of the screen, a live view of the agent operating the browser is displayed, and the instructions to the agent can be seen in the Agent Output Window.
After waiting a while, "My task is complete!" was displayed, and the process completed.
As instructed, the booking for a flight from Boston to Wolf on 2/22 was displayed.
Here is the full text of the agent's output.
You can clearly see the agent confirming whether the previous action was successful before proceeding to the next step.
🌈 Welcome to the Nova Act Playground!
💫 Agent browser connecting...
🌎 Ready to execute automated workflow.
🛝 Initializing agent with provided instructions...
🚀 Loading and analyzing target webpage...
💭 I am on the NextDot homepage. My task is to find flights from Boston to Wolf on Feb 22nd. I see a section to book a flight. The origin field is empty, but I need to set it to Boston. I should click the origin dropdown to expand the options.
💭 The origin dropdown is now expanded, meaning that my previous action was successful. I see the Boston AGI Lab option, but it is not selected. I should click the Boston AGI Lab option to select it as the origin.
💭 The origin is now set to Boston AGI Lab, so my last action was successful. I now need to set the destination to Wolf. I see the destination dropdown, but it is empty. I should click the destination dropdown to expand the options.
💭 The destination dropdown is now expanded, meaning that my previous action was successful. I see the Wolf 1061c option, but it is not selected. I should click the Wolf 1061c option to select it as the destination.
💭 The destination is now set to Wolf 1061c, so my last action was successful. I now need to set the departure date to Feb 22nd. I see the departure date field, but it is empty. I should click the departure date field to open the calendar.
💭 The calendar is now open, meaning that my previous action was successful. The calendar currently shows December 2025, but I need to set the date to February 22nd. I see a right arrow icon to navigate to the next month. I should click the right arrow icon to navigate to the next month in the calendar.
💭 The calendar now shows January 2026, meaning that my previous action was successful. I need to set the date to February 22nd, but the calendar is currently showing January. I should click the right arrow icon to navigate to the next month in the calendar.
💭 The calendar now shows February 2026, meaning that my previous action was successful. I now need to set the date to February 22nd. I see the 22 option in the calendar, but it is not selected. I should click the 22 option in the calendar to select it as the departure date.
💭 The departure date is now set to February 22, 2026, meaning that my previous action was successful. I have set the correct origin, destination, and date, but I have not searched for flights yet. I see a Search Flights button. I should click the Search Flights button to complete the search.
💭 I am now on the results page for flights from Boston to Wolf on Feb 22, meaning that my previous action was successful. I have successfully found flights from Boston to Wolf on Feb 22nd. My task is complete, and no return information was asked from me. I will simply return.
✅ My task is complete!
When you press the download button, you can download the Python code (Hello_World!.py) that was executed in the Playground.
It seems to be constructed with very simple code using the Nova Act SDK.
from nova_act import NovaAct
import os
# Browser args enables browser debugging on port 9222.
os.environ["NOVA_ACT_BROWSER_ARGS"] = "--remote-debugging-port=9222"
# Get your API key from https://nova.amazon.com/act
# Set API Key using Set API Key command (CMD/Ctrl+Shift+P) or set it below.
# os.environ["NOVA_ACT_API_KEY"] = "<YOUR_API_KEY>"
# Initialize Nova Act with your starting page.
nova = NovaAct(starting_page="https://nova.amazon.com/act/gym/next-dot/search", headless=True, tty=False)
# Running nova.start will launch a new browser instance.
# Only one nova.start() call is needed per Nova Act session.
nova.start()
# To learn about the difference between nova.act and nova.act_get visit
# https://github.com/aws/nova-act?tab=readme-ov-file#extracting-information-from-a-web-page
nova.act("Find flights from Boston to Wolf on Feb 22nd. ")
# Leaving nova.stop() commented keeps NovaAct session running.
# To stop a NovaAct instance uncomment nova.stop() - note this also shuts down the browser instantiated by NovaAct so subsequent nova.act() calls will fail.
# nova.stop()
Now that I have a general understanding of how it works, this concludes my testing.
Next Steps
Since I haven't tested the following steps, I'll introduce them based on images from the AWS official blog.
After installing the Nova Act extension in supported IDEs such as VS Code or Kiro, you can continue developing the exported code in the IDE.
Once development is complete, you can deploy to the AWS environment from the IDE.
It appears that the extension containerizes the workflow, pushes it to ECR, creates the necessary IAM roles and S3, and deploys to Bedrock AgentCore Runtime.
Additional Information
Available Regions
Currently, the only supported region is Northern Virginia.
Let's wait for it to become available in the Tokyo region.
Pricing
The pricing is as follows:
- $4.75 (per agent hour)
Agent time is the actual elapsed time while the agent is working. When multiple agents are running in parallel, each agent is charged separately.
Also, when implementing Human-in-the-loop, time spent waiting for human responses is excluded.
Conclusion
In this article, I introduced the GA and features of "Amazon Nova Act," which allows you to build AI agents that operate in web browsers.
According to the AWS official blog, it will be possible to build multi-agent workflows in cooperation with Strands Agents in the future.
Currently, since it's not available in the Tokyo region, production use might be difficult, but if you're interested, why not try testing it in the Northern Virginia region?
That's all.
