5 production challenges in multi-agent systems and how Deep Insight solved them

This article introduces Deep Insight, a production-grade multi-agent system built by AWS that automates complex data analysis by coordinating specialized agents to generate insights and reports from CSV data.

• Addresses five production challenges: execution flow control across agents, LLM model selection per role, reliable operations, secure code execution, and operational monitoring
• Uses Strands Agents SDK Graph and Agents-as-Tools patterns to orchestrate eight specialized agents (Coordinator, Planner, Supervisor, Coder, Validator, Reporter, Tracker)
• Deploys on Amazon Bedrock with Claude models tailored by role (Haiku for routing, Opus for planning, Sonnet for execution) to optimize cost and quality
• Isolates LLM-generated code execution in per-session ephemeral AWS Fargate containers behind Application Load Balancer in private VPC
• Provides Web UI with human-in-the-loop plan approval and Ops dashboard with DynamoDB task tracking, SNS failure alerts, and Cognito-protected admin access
• Open-sourced with three deployment options: Self-Hosted CLI for development, Web version for non-developers, and Managed AgentCore for production

Deep Insight demonstrates that production agentic AI success depends on architecture design, not just model capability, with validated patterns applicable across domains like marketing analytics and technology reconnaissance.