Building multi-agent systems often proves straightforward during pilot phases, yet moving these implementations into production environments reveals significant complexities. To address these difficulties, the AWS Solutions Architecture team developed Deep Insight, a production-grade multi-agent system designed to automate complex data analysis tasks while ensuring operational reliability.
Addressing Production Challenges
According to the AWS team, developers typically encounter five core challenges when scaling agentic systems: controlling execution flow, selecting appropriate large language models (LLMs) for specific tasks, ensuring reliable operations, securing the execution of AI-generated code, and maintaining system visibility.
Deep Insight addresses these areas by integrating several technologies. Execution flow is managed via the Strands Agents SDK, utilizing a Graph pattern to define order and a tool-based architecture that allows a supervisor agent to invoke specialized sub-agents—such as a coder, validator, or reporter—on demand. This structure enables a human-in-the-loop (HITL) approach, allowing users to review and revise analysis plans before execution.
Infrastructure and Security
For production deployments, the system leverages Amazon Bedrock and Amazon Bedrock AgentCore. To mitigate security risks associated with executing LLM-generated Python code, the architecture employs ephemeral, per-session containers running on AWS Fargate. These containers operate within a private VPC with restricted access, preventing the agents from accessing sensitive system files or other users’ data.
Operational visibility is achieved through integration with Amazon DynamoDB, Amazon Simple Notification Service (SNS), and Amazon Cognito. These components provide logging, alerting for system failures, and a secure administration dashboard for monitoring task metrics and performance.
Real-World Applications
The efficacy of this architectural design has been demonstrated through practical use cases. LG Electronics implemented a similar system for its marketing division, reducing the time required for data analysis from three days to 30 minutes. Additionally, the ‘TechRecon’ project presented at AWS re:Invent 2025 applied these orchestration principles to technology strategy analysis, further validating the approach as a general-purpose pattern for multi-agent workflows.
Deep Insight is available as an open-source project on GitHub, providing developers with options ranging from local CLI-based development to fully managed production environments using AgentCore.
Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/5-production-challenges-in-multi-agent-systems-and-how-deep-insight-solved-them/
Leave a Reply