Building agents takes days, taking them to production remains the hard part. The AWS Solutions Architecture team addressed this gap with Deep Insight, a production-grade multi-agent system that turns a CSV file and a natural-language prompt into charts and a validated. docx report with citations. The design uses eight specialized agents and a stack based on Strands Agents SDK, Amazon Bedrock, Amazon Bedrock AgentCore and AWS Fargate. For business, this matters as a tested template for moving from pilot demos to controlled data analysis.
How Deep Insight is designed and deployed
The system was presented as Part 1 of a three-part series about Context Engineering techniques and production deployment on AWS. In the demonstration, advertising data from fictional Yummy Food with a column definitions file is analyzed from a sales and marketing perspective. Deep Insight generates an analysis plan, shows it for approval, then executes it with real-time streaming before the finished report can be downloaded. The code is open sourced on GitHub with a setup guide.
Three deployment options share the same multi-agent architecture, so an agent validated in development moves to production as-is. The Self-Hosted version runs locally through the CLI and sets up in about 10 minutes with full access to agents, prompts and workflows. The Web version offers a browser UI for uploading data, reviewing plans and downloading reports without writing code. The Managed AgentCore version runs agents on AgentCore runtime inside a private virtual private cloud, with a custom Code Interpreter hosted separately on AWS Fargate.
Execution control combines two patterns from the Strands Agents SDK. The Graph pattern defines order as a directed graph: Coordinator to Planner to Plan Reviewer to Supervisor, with a conditional branch back to the Planner on revision requests. The Agents-as-Tools pattern lets the Supervisor invoke Coder, Validator, Reporter and Tracker only when needed. Each specialist works in an independent context and returns only compressed results, which keeps context small while preserving expertise.
What this means for business use of agents
Role separation is explicit across Coordinator, Planner, Plan Reviewer, Supervisor, Coder, Validator, Reporter and Tracker. The Coordinator handles greetings directly and routes complex tasks with a handoff_to_planner marker, using lightweight Claude Haiku to keep classification fast and low cost. The Planner uses Claude Opus with Extended Thinking and breaks work into 5-10 concrete steps in a checklist stored in a full_plan variable. The Plan Reviewer implements human-in-the-loop approval with a 300-second countdown for auto-approval and a MAX_PLAN_REVISIONS limit against infinite loops.
For companies adopting AI, the practical effect is a clearer path for automated analysis without rebuilding orchestration. Small teams can start with the Self-Hosted or Web version for rapid iteration and no-code review, while larger organizations gain session isolation with concurrent agents in isolated per-user environments and private VPC deployment. The inclusion of a Validator for calculations and a Reporter for the final document maps directly to sales, marketing and finance workflows where errors carry cost.
Limitations center on operational discipline rather than prompts alone. Teams still need to define how an agent reasons, when it verifies and what happens on failure, plus where code executes safely and how health is monitored. Responsible AI controls are promised in a separate solution detail because agents generate both code and report content. The marker to watch is Parts 2 and 3 on Context Engineering trade-offs with code and on AgentCore deployment operations.
