Integrating Strands Agents with Amazon Bedrock AgentCore Browser for Autonomous Web Tasks

Strands Agents with Bedrock AgentCore Browser

The Bedrock AgentCore Browser system provides AI agents with a secure, isolated sandbox to perform dynamic web interactions via Playwright. This tool allows agents to move beyond static text by connecting to remote browser sessions through WebSockets.

Why This Matters

While ideal AI models are often restricted to internal data processing, the technical reality of modern automation requires agents to navigate live, interactive web environments. By utilizing managed browser sessions in AWS us-east-1, developers can mitigate the security risks of local browser execution while enabling agents to handle JavaScript-heavy sites and visual verification via screenshots.

Key Insights

• Isolated execution via AgentCore Browser provides a secure sandbox for agents to utilize the Playwright library (Mindy Jen, 2026).
• WebSocket-based connectivity allows agents to connect to remote Chromium instances via CDP for real-time interaction.
• The AgentCoreBrowser tool enables the us.amazon.nova-pro-v1:0 model to autonomously summarize content from specific web URLs.
• Visual verification is supported through the page.screenshot function, allowing agents to capture and confirm UI states.
• The Strands Agent framework integrates with BedrockModel to orchestrate multi-step web tasks based on natural language instructions.

Working Examples

Direct, non-agentic browser interaction using Playwright and AgentCore Browser client.

import boto3
from bedrock_agentcore.tools.browser_client import browser_session
from playwright.async_api import async_playwright
region = "us-east-1"
# Connect to a managed browser session
with browser_session(region) as client:
    ws_url, headers = client.generate_ws_headers()
    async with async_playwright() as playwright:
        browser = await playwright.chromium.connect_over_cdp(endpoint_url=ws_url, headers=headers)
        page = await browser.new_page()
        # Navigate and capture a screenshot
        await page.goto("https://builder.aws.com/")
        await page.screenshot(path="screenshot.jpg")

Autonomous web interaction using a Strands Agent equipped with AgentCoreBrowser tools.

import boto3
from strands import Agent
from strands.models import BedrockModel
from strands_tools.browser import AgentCoreBrowser
# Initialize the browser tool
agentcore_browser = AgentCoreBrowser(region="us-east-1")
# Create an agent with browsing capabilities
agent = Agent(
    model=BedrockModel(model_id="us.amazon.nova-pro-v1:0"),
    tools=[agentcore_browser.browser],
)
# Request the agent to perform a multi-step web task
query = ("Go to https://builder.aws.com/learn/topics/amazon-bedrock-agentcore, "
         "find the first article link, and summarize its content.")
response = await agent.invoke_async(query)
print(response.message["content"][0]["text"])

Practical Applications

• Automated Quality Assurance: Using Playwright to navigate AWS Builder pages to verify element visibility; failing to handle dynamic loading can lead to execution timeouts.
• Real-time Data Collection: Strands Agents autonomously browse and summarize technical articles; hard-coding CSS selectors is an anti-pattern that causes scripts to break during UI updates.
• Complex Business Process Automation: Agents filling out forms programmatically in isolated sandboxes; neglecting to monitor WebSocket connection stability can result in lost session states.

References:

• https://dev.to/mindy_jen_phd/get-your-hands-dirty-agentcore-browser-10ck