A supply chain manager stares at a dashboard where a bright red indicator signals a critical shipment delay. The dashboard is perfect at identifying that something is wrong, but it is useless at fixing it. To resolve the crisis, the manager must now leave the dashboard and enter a manual loop of chaos, cross-referencing purchase orders in Excel, checking inventory levels in an ERP, emailing logistics partners, and reviewing contract rules in a PDF. This gap between seeing a problem and executing a solution is where most enterprise productivity dies, trapped in the friction of context switching between a dozen disconnected tools.
The Architecture of Integrated Action
Amazon Quick and the NVIDIA NeMo Agent Toolkit aim to collapse this gap by transforming the static dashboard into an active workspace. Amazon Quick serves as the conversational frontend, acting as an interactive workspace that unifies structured data with unstructured corporate knowledge. It does not merely pull from Amazon S3; it integrates knowledge sources from Google Drive, Microsoft SharePoint, Atlassian Confluence, and internal web content. To move from insight to action, Amazon Quick provides over 100 pre-built action connectors for tools like Microsoft Outlook, Slack, Jira, and Asana. This allows an analyst to identify a risk on a dashboard and immediately trigger a collaboration request or create a support ticket without ever leaving the interface.
Under the hood, the execution logic is driven by the NVIDIA NeMo Agent Toolkit. This open-source library is designed for the connection, evaluation, profiling, and optimization of agent workflows. It is built for interoperability, meaning it functions alongside existing frameworks such as LangChain, LlamaIndex, CrewAI, Microsoft Semantic Kernel, and Google ADK, or even custom Python-based agents. Developers use this toolkit to measure the precision of tool calls and tune the sequence of operations to meet specific business requirements, ensuring the agent does not just guess the next step but follows a validated logic path.
To manage the flow of information, the system employs a dual-track process. The diagnostic flow handles the initial problem identification. When a user asks about a situation, Amazon Quick analyzes dashboard metrics and knowledge sources stored in Amazon S3 to provide an answer. This automates the data gathering phase, removing the need for manual research. Once the situation is understood, the system switches to the action flow. This is where the Model Context Protocol (MCP) becomes critical. MCP provides a standardized interface that allows AI models to access external tools and data consistently.
An action request originates in Amazon Quick, passes through the Amazon Bedrock AgentCore Gateway (the MCP gateway), and reaches the AgentCore Runtime, which is a containerized hosting environment. Within this runtime, the NVIDIA NeMo Agent Toolkit workflow is executed. While Amazon Quick handles the user's intent at the frontend, the NeMo Agent Toolkit performs the heavy lifting of diagnosing supply chain risks and formulating mitigation plans. This journey from gateway to runtime container ensures that security authentication and environment controls are maintained throughout the process.
From Autonomous Guesswork to Explicit Orchestration
The fundamental shift in this architecture is the move away from total reliance on the autonomous judgment of a Large Language Model. Traditional AI agents often suffer from hallucinations, calling the wrong tool or inventing parameters because they are operating in a black box. The NVIDIA NeMo Agent Toolkit solves this by moving the workflow into an explicit design space, giving developers granular control through a three-layer hierarchy: registered functions, workflow configuration files, and the orchestrator.
Registered functions are the smallest reusable units of capability. In a supply chain context, these functions might include checking purchase order risks, analyzing inventory exposure, assessing customer impact, querying policy documents, evaluating logistics options, or generating mitigation recommendations. These functions are decoupled from the logic of when to use them. The workflow configuration file acts as the blueprint, assigning names to these functions and linking them into a specific decision-making sequence. Because the logic resides in a configuration file rather than hard-coded scripts, developers can adjust the sequence of tool calls and modify business logic without rewriting the core codebase.
The orchestrator is the engine that receives the user request and executes the functions in the order defined by the configuration file. The output of this process is not a simple text response but a comprehensive mitigation plan. This plan includes evidence, a trace of the execution path, latency metrics, and evaluator metadata. By utilizing telemetry, teams can pinpoint exactly where a bottleneck occurs in the workflow and tune the orchestration logic to increase system reliability. This structure allows a practitioner to take a sample function based on a CSV file, such as the `supply_chain_risk_orchestrator`, and replace it with a live Amazon Athena query, an RDS connection, or an ERP API call without disrupting the rest of the system.
This separation of concerns provides distinct advantages for different stakeholders. For operational leaders and analysts, the benefit is the elimination of the physical time spent between data verification and response. Instead of manually auditing supplier delays, they receive a prioritized mitigation plan directly within the context of their Amazon Quick dashboard. For engineering leaders, the benefit is a streamlined development cycle. Because Amazon Quick handles the UI and NVIDIA NeMo handles the logic, developers can stop focusing on frontend tweaks and spend their time optimizing tool-call accuracy and backend performance.
Deploying the Agentic Workflow
Implementing this workflow requires a specific infrastructure setup. Practitioners must first install the AWS CLI and configure their IAM Identity Center or SSO profiles. Authentication is handled via the terminal using the following command:
aws sso loginOnce authenticated, the infrastructure is deployed using AWS CloudFormation templates and deployment scripts. These scripts automate the configuration of the Amazon Quick dashboard, the connection of knowledge sources, the setup of the chat agent, and the linking of MCP actions.
The runtime implementation lives inside the Amazon Bedrock AgentCore Runtime container. When Amazon Quick triggers an MCP action, the AgentCore Gateway routes the request through a Lambda target adapter to the runtime endpoint. Inside the container, a FastAPI application listens for requests at the `/invocations` endpoint. Upon receiving a request, it loads the NeMo workflow, and the orchestrator executes the registered functions to derive the final mitigation plan. This containerized approach ensures that business logic is isolated, meeting enterprise security requirements while allowing for flexible scaling.
For those moving from a pilot to production, the strategy is a phased replacement of data sources. The initial deployment typically uses CSV-based functions to validate the logical flow of the agent. Once the workflow is proven, these are replaced with real-time data connectors such as Amazon Athena, RDS, or external WMS and TMS APIs. This approach reduces development risk by verifying the agent's reasoning before connecting it to mission-critical production data. The complete implementation code and templates are available at https://github.com/aws-samples/amazon-quick-nemo-agent-toolkit.
Moving beyond the novelty of chatting with a bot requires a shift toward rigorous workflow design. The integration of MCP and the three-layer architecture of NVIDIA NeMo transforms the AI from a conversationalist into a functional employee capable of executing complex business processes. The performance of an agent is no longer measured by the fluency of its prose, but by the precision of its orchestration from data discovery to resolution.
