IBM watsonx.ai Use Case

The Conversational AI Agent centers around a base agent architecture and interacts with IBM Maximo systems through natural language. Each agent has a specific name, defined purpose, prompt templates, access to external tools, and is powered by a Large Language Model(LLM) via LangGraph, LangChain and watsonx.ai.
‍

This setup enables agents to process a user's natural language inputs through a structured workflow, allowing them to retrieve Maximo data and perform actions like creating new work orders or initiating workflows. The image below shows an example of a user requesting work orders with priority 1 on December 31, 1998. The Supervisor Agent routes the request to the Maximo Agent, which queries the Maximo API. Results are returned and formatted into both natural language and JSON for the user to review as shown in Figure 1.
‍

‍

Agent Structure

The use case implementation uses a multi-agent system built using LangGraph, LangChain and watsonx.ai with several specialized agents. Each base agent is configured with the following attributes:

• Name: Identifier for logging and agent selection

• Purpose: Task or domain-specific focus (e.g., retrieving Maximo records)

• Prompts: Instructional templates to guide the LLM’s responses

• LLM: Connected watsonx.ai model that generates natural language output

• Tools: APIs or utility modules that the agent can call to perform actions
  ‍

The agent is also designed with core functions such as:

• Handle Input: Manages user requests and interprets context

• Tool invocation and response routing based on agent logic

‍

Agent Flow Architecture

‍

1. Input: The system starts with the Supervisor Agent which acts as the main coordinator
2. Agent Types
   
   1. Supervisor Agent: Main Coordinator
      
      1. Receives user input
      2. Decides which specialized agent should handle the request
      3. Generates final response
   2. Maximo Agent: Handles Maximo-specific queries and operations
      
      1. Processes Maximo-related requests
      2. Uses Maximo tools for specific operations
      3. Returns results to the Supervisor Agent
   3. Vector DB Agent: Handles vector database operations
      
      1. Processes queries related to vector databases
      2. Uses vector database tools for operations
      3. Returns results to the Supervisor Agent
3. Flow Control
   
   1. The Supervisor Agent uses a router function to determine which specialized agent should handle the request
   2. Each specialized agents have their own functions to determine if it needs to use its specific tools
   3. After tool usage, the control returns to the respective agent
   4. Results flow back to the Supervisor Agent for final processing
4. State Management
   
   1. The system maintains states through the AgentState class
   2. States include:
      
      1. User Input: Original request provided by the user
      2. Supervisor Decisions: The logic used by the Supervisor to determine which specialized agent should handle the task
      3. Maximo Payload: Structured data sent to or received from Maximo APIs
      4. Tool Calls: The execution of external functions or APIs triggered by an agent to complete a task
      5. Agent Responses: The intermediate outputs or messages generated by the specialized agents
      6. Final Response: The completed, user-facing answer by the Supervisor
      7. Memory Chain: Stored context of previous interactions to maintain continuity
5. Exit Point: The Supervisor Agent is also the finish point as it generates the final response based on the specialized agents’ outputs

‍

The architecture allows for modular handling of different types of requests, has clear separations of functionalities between different agents, implements flexible routing of requests based on their nature, and centralized coordination through the Supervisor Agent.

‍

Exploring Use Case Ideas

In the demo with IBM, no complete use case was implemented beyond the one described in this document. However, the following ideas were discussed:

• Work Order Retrieval: Users can ask the agent about specific work orders in Maximo, and the agent can fetch and summarize relevant details through API calls
• Work Order Creation/Workflow Execution: The agent could guide users through the process of creating a new work order or triggering predefined workflows, using natural language input
• Visual Inspection Support: Possibility of Integrating with Maximo Visual Inspection (MVI) and mobile tools to assess asset damage through images or videos

‍

This use case highlights how conversational AI can make working with Maximo much easier and more intuitive. By using natural language, users can ask the agent to find or manage work orders without needing to understand complex systems or technical processes. Behind the scenes, the agent is built with different ‘helper agents’, each with a specific role, working together to understand the request, fetch data from Maximo, and return clear, concise responses.  The Supervisor Agent acts like a team lead, making sure the right agent handles the task and then organizing the final reply.

‍

The current example focuses on retrieving work orders based on specific criteria, but the same setup can be expanded to do much more – like creating work orders, starting workflows, or even analyzing images from Visual Inspection. This approach shows how advanced AI tools like watsonx.ai can simplify complex workflows and help teams interact with enterprise systems in a more natural and efficient way.