Build specialized agent workflows for your business with Amazon Quick and NVIDIA NeMo Agent Toolkit

Supply chain teams have a familiar problem: plenty of dashboards telling them something's wrong, and almost no fast path to figuring out what to actually do about it. A late shipment from a supplier can trigger a scavenger hunt through purchase orders, inventory levels, customer contracts, freight rules, and approval chains before anyone can recommend a fix. AWS's new reference architecture, built with NVIDIA's open-source NeMo Agent Toolkit, is an attempt to close that gap by turning Amazon Quick into the front door for a full agentic investigation.

The setup splits work cleanly. Amazon Quick handles the parts business users actually see — dashboards, chat, knowledge search across sources like S3, SharePoint, and Confluence, plus more than 100 prebuilt connectors into tools like Jira, Slack, and Outlook. Everything heavier lives behind the scenes in NeMo Agent Toolkit, running inside an Amazon Bedrock AgentCore Runtime container and reached through an AgentCore Gateway MCP endpoint. When an analyst asks a simple 'what's happening' question, Quick answers straight from the dashboard and its knowledge base. Ask 'what should I do about it,' though, and Quick fires off an MCP action that hands the real decision-making to NeMo's orchestrator.

That orchestrator, called supply_chain_risk_orchestrator, chains together six purpose-built tools: one flags delayed purchase orders, another calculates days of inventory supply and stockout odds, a third maps at-risk SKUs to specific customer revenue and SLA exposure, a fourth checks contract and freight-approval policy, a fifth hunts for expedite or substitution options, and a final one ranks the mitigation choices with built-in critic checks. The whole chain is defined in a plain YAML config, which means developers can swap in real data sources later — Athena queries, ERP APIs, WMS/TMS systems — without rewriting the orchestration logic itself. In the demo, everything runs off sample CSVs, but the architecture is clearly built with production replacement in mind.

What's notable here isn't the individual pieces — purchase-order tracking and inventory math aren't new — but the packaging. AWS is pitching this squarely at teams that don't have the headcount to run large planning departments: fast-growing companies, startups scaling order volume, supply-chain leads who need repeatable decisions rather than one-off spreadsheet detective work. NeMo Agent Toolkit's observability layer, tracking per-step latency and evaluation metrics, is the part that signals AWS and NVIDIA expect this to graduate from demo to production, where knowing why an agent made a call matters as much as the call itself.

The whole thing ships as a CloudFormation-deployable sample repo, with CodeBuild handling the container build so nobody needs Docker installed locally. It's a fairly involved setup — SSO auth, IAM roles, ECR repos, Lambda targets — for what amounts to a supply-chain chatbot with better plumbing. But that plumbing, connecting a business-friendly BI tool to a framework-agnostic agent backend via MCP, is the actual template AWS wants people to copy for other domains entirely.