Knowledge Base Academy Community My Tickets
Fabric 8.4
Fabric 8.4

Agent Framework Components: Agents, Tools and Utilities

An AI agent is a module that receives data, context, and resources, then uses an AI model for reasoning, planning, and decision-making to achieve specific goals.

Agent Types

The GenAI Data Fusion framework includes two categories of agents, both represented as Broadway actors and flows.

Built-in Agents

Workflow Orchestrating Agents

Agent Role
Orchestrator Manages the overall agentic flow and coordination
Reflector Evaluates the user's request and selects the appropriate response path
Refiner When sub-agent path is chosen, it refines the user's request into a concise, actionable goal for a selected sub-agent
Planner Builds and executes multi-step task plans

Data Agents

Agent Role
Data Retrieve Explores Fabric Data Products by generating SQL dynamically

Project Implementation Agents — Worker Sub-Agents

The implementation agents act as worker sub-agents in the agentic workflow, with the aim of accomplishing domain-specific goals. These are Broadway flows tagged to handle specific domains or request types.

Example: The loans_subagent tag identifies an agent specialized in banking loan inquiries.

Read more practice information here about the agentic workflow's built-in agents and worker sub-agents.

Tools

Tools are Broadway flows designed to perform specific actions.

When to Use Tools

While the generic Data Retrieve agent handles most queries using schema descriptions, tools become essential for:

Scenario Reason
Sophisticated queries Complex business rules requiring predefined SQL statements crafted by data specialists
Actions with business impact Controlled flows for operations involving monetization, compliance, or auditing requirements
External system integration Dedicated calls to services outside Fabric's responsibility

Tool Identification

Tools are identified by:

  • Tags — one or more tags for discoverability
  • Description — a clear explanation of the tool's purpose
  • Parameter Remarks — documentation on input/output parameters

This information is passed to the LLM to aid in tool selection during plan execution.

Tool Types

While there are several built-in tools like queryExecute, tools are usually built as part of the project Implementation.

It is recommended to build tools at the data products. At your agentic flow you shall specify these tools, where AI Fusion agentic flow can call them either directly - as Broadway flow - from aifusion or from the relevant the data product, or as MCP.

Read here for more information about tools calling and consumption, according to tool tags naming conventions.

Note: Tools typically do not involve AI unless an agent is tagged as a tool.

Utility Actors

The AI Fusion platform uses several utility actors for working with AI, such as LLMConst, LLAppend and LLInvoke.

For more information, read here.

Agent Framework Components: Agents, Tools and Utilities

An AI agent is a module that receives data, context, and resources, then uses an AI model for reasoning, planning, and decision-making to achieve specific goals.

Agent Types

The GenAI Data Fusion framework includes two categories of agents, both represented as Broadway actors and flows.

Built-in Agents

Workflow Orchestrating Agents

Agent Role
Orchestrator Manages the overall agentic flow and coordination
Reflector Evaluates the user's request and selects the appropriate response path
Refiner When sub-agent path is chosen, it refines the user's request into a concise, actionable goal for a selected sub-agent
Planner Builds and executes multi-step task plans

Data Agents

Agent Role
Data Retrieve Explores Fabric Data Products by generating SQL dynamically

Project Implementation Agents — Worker Sub-Agents

The implementation agents act as worker sub-agents in the agentic workflow, with the aim of accomplishing domain-specific goals. These are Broadway flows tagged to handle specific domains or request types.

Example: The loans_subagent tag identifies an agent specialized in banking loan inquiries.

Read more practice information here about the agentic workflow's built-in agents and worker sub-agents.

Tools

Tools are Broadway flows designed to perform specific actions.

When to Use Tools

While the generic Data Retrieve agent handles most queries using schema descriptions, tools become essential for:

Scenario Reason
Sophisticated queries Complex business rules requiring predefined SQL statements crafted by data specialists
Actions with business impact Controlled flows for operations involving monetization, compliance, or auditing requirements
External system integration Dedicated calls to services outside Fabric's responsibility

Tool Identification

Tools are identified by:

  • Tags — one or more tags for discoverability
  • Description — a clear explanation of the tool's purpose
  • Parameter Remarks — documentation on input/output parameters

This information is passed to the LLM to aid in tool selection during plan execution.

Tool Types

While there are several built-in tools like queryExecute, tools are usually built as part of the project Implementation.

It is recommended to build tools at the data products. At your agentic flow you shall specify these tools, where AI Fusion agentic flow can call them either directly - as Broadway flow - from aifusion or from the relevant the data product, or as MCP.

Read here for more information about tools calling and consumption, according to tool tags naming conventions.

Note: Tools typically do not involve AI unless an agent is tagged as a tool.

Utility Actors

The AI Fusion platform uses several utility actors for working with AI, such as LLMConst, LLAppend and LLInvoke.

For more information, read here.

© Copyright 2026 K2VIEW