Kyungshin Holdings Builds Internal AI Agent Platform Based on Amazon Bedrock AgentCore

Kyungshin Holdings built a platform that securely connects internal data and centrally manages agents using Amazon Bedrock AgentCore. The initiative focused on addressing three core requirements for the organization's AI Transformation (AX): data connectivity, execution environment, and asset sharing.

Secure MCP Tooling of Internal Data

Existing MS-SQL and SAP systems were exposed as MCP tools via AgentCore Gateway without modification. Authentication credentials are managed in AWS Secrets Manager, and VPC private endpoints are used to securely access the internal network without internet exposure. The Text-to-SQL agent applies read-only permissions and a SQL parser (sqlglot) to block DDL/DML operations and maintain audit logs.

Performance Improvement and Operational Efficiency

The architecture was optimized from an initial multi-agent structure to a combination of a single agent and Skills, improving execution time by approximately 3x to 5x. By introducing Skills documentation authored by domain experts, the number of tool calls for the same question was reduced from 8 to 1. AgentCore Runtime isolates user sessions in microVMs, while AgentCore Memory automatically manages conversation history.

No-code Agent Creation and Governance

A workflow for searching and approving agents, MCP servers, and skills is operated through AWS Agent Registry. Non-developer employees can create specialized agents by combining models, tools, and skills without code via AgentCore Harness. Usage and costs per user are tracked via the Insights dashboard, and agent performance is periodically evaluated and optimized using AgentCore Evaluations.