AI agent definition
An AI agent uses a large language model (LLM) as its cognitive engine to interpret natural-language requests, break down complex goals, and make autonomous decisions.
Because a large language model alone cannot interact with external enterprise systems, the AI agent relies on predefined capabilities known as tools. By selecting and running the right tool for each request, the AI agent bridges the gap between a conversational prompt and a real-world action.
In HCL Universal Orchestrator, workflows are exposed to the AI agent as these tools, instead of granting it direct, lower-level API access. Running a pre-orchestrated workflow takes fewer reasoning steps from the LLM than assembling a complex IT process from individual API calls. This approach also enforces governance. The AI agent can trigger only the workflows added to its definition, and can read only the output properties each workflow explicitly defines. Other enterprise data stays hidden from the model.
You build an AI agent from the Blocks tab of the Graphical Designer or from the Orchestration CLI, and store the definition in a folder like any other HCL Universal Orchestrator item.
Definition components
An AI agent definition specifies the following elements:
- Agent type
- Determines which AI model layer powers the agent: Basic, External MCP, or Agentic AI Builder. For a comparison of the three types, see Managing agent types in the AI agent.
- Workflows
- The set of workflows that the agent can select and run to satisfy a request. Each workflow must define its service name, description, and variables so the agent knows what input to pass. The agent always knows whether a workflow run completed, so output properties are not mandatory when the agent only needs to confirm completion before moving on. Define output properties when the agent needs specific data back from the run; without them, the agent reads no other result. Publishing the workflow on the Self-Service Catalog is not required; adding it to this list is enough. For details, see Creating workflows as a service using Orchestration CLI and Output properties and data transfer between workflows. When the agent runs as a task, the invoked workflow also receives metadata about the originating task and its parent workflow; see AI agent task context variables.
- Chatbot interface
- An optional conversational entry point, published in the AI Agents section of the UI, that lets users interact with the agent directly instead of running it as a task.
For the full list of fields and their configuration, see Managing an AI agent.
Interacting with an agent
After you create an AI agent, you can trigger it as a task from a workflow, or, if the chatbot interface is enabled, converse with it directly. For details, see Communicating with the AI agent.
Connecting to an external MCP server
If the agent type is External MCP or Agentic AI Builder, the agent exposes a generated server URL that lets an external client, or an Agentic AI Builder package, call HCL Universal Orchestrator resources directly. For the steps required to establish this connection, see Course of action if the agent type is External MCP.