Samsung Account Builds Multi-Agent AIOps System on Amazon Bedrock AgentCore

Samsung Account is a large-scale unified authentication system serving approximately 2.1 billion users worldwide, operated with a multi-region active-active architecture. As the service scale expanded, the operational complexity for tasks such as root cause analysis and security audits increased exponentially due to the interplay of multi-region deployments, multiple accounts, and numerous microservices.

In response, the Samsung Account SRE team adopted an Agentic AIOps system based on Amazon Bedrock AgentCore. To mitigate operational risks stemming from LLM non-determinism, they adopted a phased trust-building approach under the principle of "agents analyze, humans approve execution," rather than fully autonomous operations.

Hierarchical Multi-Agent Architecture

The system is designed with a three-tier structure consisting of an Orchestrator, domain Supervisors, and Sub-agents to overcome the limitations of a single agent. The top-level Orchestrator analyzes requests and routes them to the appropriate domain Supervisor, which then invokes Sub-agents to perform actual tasks within six operational domains, such as incident analysis, security audits, and cost optimization.

Design for Convergence and Scalability

Requests entering through two entry points, Slack and the web console, converge into the same Orchestrator, ensuring consistent analysis logic and guardrails. Additionally, by modularizing common functions such as model invocation, memory, and account access, and managing agents via a registry, the system ensures scalability while maintaining overall consistency when adding new agents.