Building AI agents for business support using Amazon Bedrock AgentCore

Background

Works Human Intelligence (WHI), a provider of integrated HR systems for major Japanese organizations, faced operational challenges from routine tasks like processing commuting allowance approvals and executing manual browser-based HR system checks. To address this, WHI collaborated with the AWS Generative AI Innovation Center (GenAIIC) to develop two specialized AI agents using Amazon Bedrock AgentCore, aiming to automate these processes and reduce workload.

Key Points

1. Commuting Allowance Agent: Architectural Evolution for Flexibility

• Initial Challenge: The agent was initially built as a monolithic application using LangGraph on Amazon ECS and Fargate. This configuration raised concerns about scalability and flexibility for future expansion.
• Core Solution: The architecture was redesigned to leverage Amazon Bedrock AgentCore Runtime. Sub-agents were decoupled from the main application and run individually on separate Runtimes. This shift facilitates easier future expansion and integration of new sub-agents.
• Multi-Tenancy & Security: To support multiple corporate customers, the solution uses Amazon DynamoDB to manage tenant data and Amazon Cognito for authentication and authorization, maintaining WHI's control over customer management.

2. Browser Operation Agent: Automating Manual Workflows

• Core Function: This agent performs tasks within the "COMPANY" HR system on behalf of users, such as checking content, performing operations, and collecting evidence. This directly addresses the manual, repetitive work of operational departments.
• Orchestration & Efficiency: The agent orchestrates browser actions to complete these tasks, fundamentally shifting the role of human operators from execution to oversight and exception handling.

3. Unified Platform and Observability

• Consolidated Technology: Both agents were built on Amazon Bedrock AgentCore, providing a unified platform for development, deployment, and management.
• Reduced Operational Overhead: Previously, WHI hosted and managed Langfuse for agent monitoring, incurring operational costs. Migrating to AgentCore Observability eliminated this burden, offering built-in monitoring capabilities and reducing complexity.

Significance

This project demonstrates a practical blueprint for deploying AI agents to solve specific, high-volume operational problems in enterprise settings. The key innovations are:

• Cost-Effective Architecture: The move to a distributed agent runtime model directly contributed to the 97% cost reduction, showcasing how intelligent architectural choices can dramatically improve the ROI of AI implementations.
• Tangible Automation: The agents address concrete pain points—approval workflows and system navigation—freeing human employees to focus on more complex, value-added work.
• Scalable Foundation: The use of a managed service like AgentCore provides a scalable and maintainable foundation, allowing WHI to evolve its AI agents alongside customer needs without heavy infrastructure management.