Build Stateful IT Service Desk Agent with LangGraph on EKS
This post demonstrates building an AI-powered IT service desk agent using LangGraph on Amazon EKS that can autonomously handle routine L1 support requests and escalate complex issues to human engineers. The agent leverages LangGraph's stateful workflow capabilities, such as interrupt() and checkpointing, with Amazon DynamoDB for state persistence, providing context-preserving escalations and durable state across failures. The solution uses Amazon OpenSearch Serverless for retrieval, Anthropic's Claude on Amazon Bedrock for generation, and Karpenter for efficient node autoscaling, offering a portable and scalable pattern for IT support automation.
- →LangGraph for Tiered Support Workflows
- →Durable State Persistence and Escalation
- →Hybrid Confidence Routing for Escalation
- →IAM Roles for Service Accounts (IRSA)
- →Stateful AI Agent for IT Support Automation
Security (1) ›
- IAM Roles for Service Accounts (IRSA)
The architecture utilizes IRSA to grant pods necessary permissions to access AWS services like Amazon Bedrock, DynamoDB, and OpenSearch Serverless without relying on static credentials.
Features (3) ›
- LangGraph for Tiered Support Workflows
LangGraph's graph-based modeling, interrupt(), and checkpointing primitives are employed to manage tiered support escalations, ensuring context is preserved during human intervention.
- Durable State Persistence and Escalation
Amazon DynamoDB is used for durable state persistence, allowing escalated tickets to remain intact across pod restarts, scaling events, and multi-replica routing, with OpenTelemetry providing execution tracing.
- Hybrid Confidence Routing for Escalation
The agent uses a hybrid approach combining retrieval scores from Amazon OpenSearch Serverless and LLM self-assessed confidence to determine when to escalate to human engineers, addressing knowledge gaps or ambiguous contexts.
Enhancements (1) ›
- Optimized Architecture with AWS Services
The solution integrates Amazon EKS for orchestration, Amazon OpenSearch Serverless with Amazon Titan embeddings for semantic search, Anthropic's Claude on Amazon Bedrock for generation, and Karpenter with Spot Instances for cost-efficient autoscaling.
Notes (1) ›
- Stateful AI Agent for IT Support Automation
An AI agent is presented that autonomously handles routine IT support requests and escalates complex issues to human engineers, built using LangGraph for stateful workflows and deployed on Amazon EKS.
https://aws.amazon.com/blogs/opensource/building-a-stateful-it-service-desk-agent-with-langgraph-on-amazon-eks/
Related releases
- Amazon EKS adds Kubernetes version rollback feature AWS News Blog ·
- Amazon EKS Auto Mode adds Application Recovery Controller zonal shift AWS What's New ·
- Amazon EMR on EKS adds Apache Spark troubleshooting agent AWS What's New ·
- AWS Config adds 191 managed rules for AI and core services AWS What's New ·
- AWS Neuron 2.31.0 enhances performance and adds EKS operator AWS What's New ·
- Amazon ECS Managed Instances reduce GPU management fees AWS What's New ·