Hyundai AutoEver's Multi-Agent AI on Amazon Bedrock: Cutting Incident Response Time to 5 Minutes

Hyundai AutoEver's Vehicle Control Service Development Team built ErrorWatcher at an AWS GenAI Sandbox hackathon, cutting incident response time from hours to 5 minutes. The team won the Spotlight Award by solving a real operational problem with generative AI.

The core is multi-agent orchestration based on LangGraph. The system has four agents—Monitor, Detective, Solver, and Reporter—that work together sequentially to automate symptom collection, root cause analysis, solution proposal, and report generation.

The architecture is as follows.

• LLM: Claude Sonnet on Amazon Bedrock
• Orchestration: LangGraph on AWS Lambda
• RAG: Amazon Bedrock Knowledge Base and OpenSearch Serverless
• Storage: Amazon S3
• UI: Streamlit

Instead of the fully managed Amazon Bedrock Agents, the team chose custom orchestration based on StateGraph to finely control state handoff and execution flow between agents. They also applied metadata filtering to Knowledge Base retrieval, configuring Detective to reference only past-incidents/ and architecture/, Solver to reference only runbooks/, and Reporter to reference only templates/.

Continuous learning was also built in. Analysis results are assigned HIGH/MEDIUM/LOW confidence levels, and high-confidence reports are automatically saved to learned-cases/ to serve as a basis for analyzing similar incidents later. For example, in an EKS Pod CrashLoopBackOff situation, the system narrows the cause down to OOMKilled due to insufficient memory limits, and suggests immediate actions like kubectl patch and rollout restart along with long-term improvements such as introducing HPA.

Going forward, the team plans to develop Proactive Monitoring that watches key metrics in advance, Human-in-the-Loop automation combining Teams alerts with an approval process, and Teams Webhook integration enabling chat-based requests.