AWS Launches Bedrock AgentCore Browser Tool for AI-Driven Legacy Web Automation
Summary
Updated Aug 14 — AWS publishes tutorial integrating SageMaker AI endpoints with AgentCore Runtime, enabling mixed-model multi-agent systems combining Bedrock and self-hosted models.
- • AWS released Bedrock AgentCore Browser Tool, a managed service enabling AI agents to automate legacy enterprise web applications.
- • Uses Playwright and Chrome DevTools Protocol (CDP) to give agents human-like control of any HTTP/HTTPS app via isolated Chromium sessions.
- • Integrates with Strands Agents for multi-agent orchestration and Amazon Bedrock foundation models with AWS IAM security and full audit trails.
- • Targets industries like insurance, healthcare, and finance where manual data entry into legacy systems remains a major cost and compliance burden.
Updates
SageMaker AI + AgentCore Runtime: Mixed-Model Multi-Agent Systems
AWS demonstrates combining OpenAI-compatible SageMaker AI endpoints with Bedrock AgentCore Runtime, allowing specialized agents to use cost-optimized or domain-specific models alongside Bedrock foundation models in a single production-ready architecture with data residency control.
Reference Architecture: Orchestrator + Budget + Financial Analysis Agents
Published reference implementation uses three agents unified through AgentCore Runtime: Claude Haiku 4.5 as orchestrator on Bedrock for intent routing, Claude Sonnet 4.6 for budget breakdowns on Bedrock, and Qwen 3.5 9B on SageMaker AI for stock analysis and portfolio construction via tool-calling.
OpenAI-Compatible SageMaker Endpoints Enable Framework Integration
SageMaker AI endpoints expose an OpenAI-compatible API, letting orchestration frameworks like Strands Agents route tasks to custom or third-party models on SageMaker without modifying agent logic — including token-level observability not available from Strands by default. Full source code on GitHub.
Details
Bedrock AgentCore Browser Tool Released
Fully managed, cloud-based browser service that lets AI agents interact with legacy web applications through secure, isolated browser sessions via Playwright and WebSocket-based Chrome DevTools Protocol (CDP).
Playwright and CDP Architecture
Runs a managed Chromium instance in the cloud; connects to legacy apps via WebSocket-based CDP. Targets any app accessible over HTTP or HTTPS regardless of original browser design.
Strands Agents Integration for Orchestration
Combined with Strands Agents for model-driven orchestration, enabling workflows ranging from simple single-step automations to complex multi-agent pipelines powered by Amazon Bedrock foundation models.
Session-Isolated IAM Security and Audit Trails
Each browser session is isolated with AWS IAM controls and full audit trails, supporting regulatory compliance requirements in healthcare, insurance, and financial services.
Legacy Enterprise Systems Remain Automation Bottleneck
Enterprises in healthcare, manufacturing, retail, and financial services face significant costs from manual data entry into legacy HTML systems. Insurance companies, for example, process tens of thousands of policy changes annually via multi-step web workflows.
Open-Source Terraform Reference Implementation on GitHub
Complete source code and Terraform deployment blueprint available on GitHub (aws-samples/sample-agentcore-browser-legacy-automation) to accelerate enterprise adoption.
Details about AWS Bedrock AgentCore Browser Tool launch and capabilities.
What This Means
AWS's Bedrock AgentCore Browser Tool addresses a persistent enterprise automation gap: legacy web applications with server-rendered HTML that can't be reached by APIs or standard RPA. By deploying AI agents with full browser control in isolated cloud sessions, enterprises can finally automate complex multi-step workflows in systems like insurance policy platforms without replacing them. The IAM integration and audit trail support make this viable for regulated industries. This is part of AWS's broader push to make Amazon Bedrock the enterprise platform for agentic AI workloads.