Knowledge Base Academy Community My Tickets
Fabric 8.4
Fabric 8.4

Agentic Flow & Agents in Practice

This article provides a practical, implementation-focused guide to the agents that participate in the Agentic Flow within the AI Fusion framework.

While The Agentic Flow article explains the conceptual execution model — how decisions are made, how paths are selected, and how responsibilities are delegated — this article translates these concepts into actual agents, focusing on how to set built-in agent attributes and how to create, configure, and reference sub-agents.

The following diagram illustrates a typical agentic flow, as viewed from the perspective of AI Fusion agents and tools:

aifusion agentic flow

The Orchestrator

When the conversation entry-point flow calls the Orchestrator actor, it passes several input parameters. These inputs and attributes are then used by the Orchestrator flow itself and by its complementary agents — Reflector, Refiner and Planner. All these agents can be considered building blocks that you do not need to touch.

Input Description Example
userQuery The user's question or request "What is my credit card balance?"
synopsis Business entity story/summary "David Smith, Retail customer, 2 credit cards..."
toolMTable The name of the MTable containing the list of tool tags and their definitions.
It is used by the Planner to tighten and limit which tools can be used, while Planner agent evaluates whether and which tools to use.		 Banking_tool_tags
subagents_tag A tag used for identifying the sub-agents available for this flow.
This is used by both the Reflector, to look for and identify matching sub-agents, and by the Planner, when evaluating whether and which sub-agent to use.		 Banking_subagent
responderPrompt Guidelines for grounding the final response to the conversation caller (end user). They usually include rules and instructions for formalizing the answer. TBD
searchProceduresFlowName A flow used for retrieving corporate procedures.
The procedures are used by the Planner, where this flow is responsible to retrieve the relevant procedures to augment the Planner context prompt. Procedures are usually retrieved via a vector store search.		 Banking_searchProcedures
searchPlanFlowName A flow used for retrieving sample plans.
These samples are used by the Planner, where this flow is responsible to retrieve the relevant samples to augment the Planner context prompt, helping it to build a plan. Plans are usually retrieved via a vector store search.		 Banking_searchPlan

The Planner

If a request requires gathering information or executing multiple steps but does not match a predefined sub-agent, the orchestrator_planner is triggered.

The LLM is tasked with generating a step-by-step execution plan using several resources:

  1. Tools and Agent List: A list of available and relevant (by tags) Broadway flows, along with descriptions and remarks added to input parameters. The LLM uses these descriptions to choose the most-suitable tool and determine the required inputs. Relevant sub-agents (by tag) are also providers.
  2. Sample Plans (Optional): JSON files (e.g., Banking_plans.json) — containing pre-built templates — that show the LLM how to combine tools to accomplish similar objectives.
  3. Corporate Procedures (Optional): Documents that define business rules or provide step-by-step instructions (e.g., verification criteria for credit limit reduction).

Once the plan steps are prepared, the Planner executes them step-by-step.

Note: It is a best practice to rely on sub-agents to achieve a higher performance, as the Planner approach is slower and less predictable due to the required plan-generation step.

The Worker Sub-agents

A worker sub-agent is a tagged Broadway flow (for example, loans_subagent) designed to handle a specific domain or category of requests, such as banking loans.

Sub-agent Discovery

The Reflector agent:

  1. Searches for flows with agent tags
  2. Examines the flow descriptions
  3. Identifies the specialized agent best suited for the current task

When an appropriate sub-agent is found, the Orchestrator calls the Refiner agent to prepare the sub-agent's goal based on the user request and context.

Notes:
  1. Use the sub-agent Broadway flow properties to add tags and descriptions.
  2. Agent tags are specified as an attribute of the Orchestrator agent. Providing it with all flows tagged as sub-agents can confuse and overwhelm the agent when selecting the appropriate sub-agent.

Sub-agent Card

Input

When invoking a sub-agent, it receives detailed context, including:

  • Its Role and Objectives (as a system message).
  • Domain Data (details on relevant LUs, reference tables, column names, and descriptions) to enable dynamic SQL query construction.
  • A specific List of Tools (other Broadway flows) tailored to the domain.

Logic

A typical sub-agent is composed of logic steps alongside LLMAgent actor usage, providing improved control and reliability.

The LLMAgent actor decides which tools to activate and how to formulate responses based on:

  1. Predefined Domain Data — retrieved using a DBCommand actor
  2. Schema Information — tables relevant for SQL crafting
  3. System Prompt — contains agent goals and behavioral guidelines
  4. User Prompt — the refined user query
  5. Tool List — available tools that the AI can invoke

Use Broadway flow properties to add tags and descriptions to sub-agent flows. The framework uses these properties to select the appropriate agent for each request.

Agentic Flow & Agents in Practice

This article provides a practical, implementation-focused guide to the agents that participate in the Agentic Flow within the AI Fusion framework.

While The Agentic Flow article explains the conceptual execution model — how decisions are made, how paths are selected, and how responsibilities are delegated — this article translates these concepts into actual agents, focusing on how to set built-in agent attributes and how to create, configure, and reference sub-agents.

The following diagram illustrates a typical agentic flow, as viewed from the perspective of AI Fusion agents and tools:

aifusion agentic flow

The Orchestrator

When the conversation entry-point flow calls the Orchestrator actor, it passes several input parameters. These inputs and attributes are then used by the Orchestrator flow itself and by its complementary agents — Reflector, Refiner and Planner. All these agents can be considered building blocks that you do not need to touch.

Input Description Example
userQuery The user's question or request "What is my credit card balance?"
synopsis Business entity story/summary "David Smith, Retail customer, 2 credit cards..."
toolMTable The name of the MTable containing the list of tool tags and their definitions.
It is used by the Planner to tighten and limit which tools can be used, while Planner agent evaluates whether and which tools to use.		 Banking_tool_tags
subagents_tag A tag used for identifying the sub-agents available for this flow.
This is used by both the Reflector, to look for and identify matching sub-agents, and by the Planner, when evaluating whether and which sub-agent to use.		 Banking_subagent
responderPrompt Guidelines for grounding the final response to the conversation caller (end user). They usually include rules and instructions for formalizing the answer. TBD
searchProceduresFlowName A flow used for retrieving corporate procedures.
The procedures are used by the Planner, where this flow is responsible to retrieve the relevant procedures to augment the Planner context prompt. Procedures are usually retrieved via a vector store search.		 Banking_searchProcedures
searchPlanFlowName A flow used for retrieving sample plans.
These samples are used by the Planner, where this flow is responsible to retrieve the relevant samples to augment the Planner context prompt, helping it to build a plan. Plans are usually retrieved via a vector store search.		 Banking_searchPlan

The Planner

If a request requires gathering information or executing multiple steps but does not match a predefined sub-agent, the orchestrator_planner is triggered.

The LLM is tasked with generating a step-by-step execution plan using several resources:

  1. Tools and Agent List: A list of available and relevant (by tags) Broadway flows, along with descriptions and remarks added to input parameters. The LLM uses these descriptions to choose the most-suitable tool and determine the required inputs. Relevant sub-agents (by tag) are also providers.
  2. Sample Plans (Optional): JSON files (e.g., Banking_plans.json) — containing pre-built templates — that show the LLM how to combine tools to accomplish similar objectives.
  3. Corporate Procedures (Optional): Documents that define business rules or provide step-by-step instructions (e.g., verification criteria for credit limit reduction).

Once the plan steps are prepared, the Planner executes them step-by-step.

Note: It is a best practice to rely on sub-agents to achieve a higher performance, as the Planner approach is slower and less predictable due to the required plan-generation step.

The Worker Sub-agents

A worker sub-agent is a tagged Broadway flow (for example, loans_subagent) designed to handle a specific domain or category of requests, such as banking loans.

Sub-agent Discovery

The Reflector agent:

  1. Searches for flows with agent tags
  2. Examines the flow descriptions
  3. Identifies the specialized agent best suited for the current task

When an appropriate sub-agent is found, the Orchestrator calls the Refiner agent to prepare the sub-agent's goal based on the user request and context.

Notes:
  1. Use the sub-agent Broadway flow properties to add tags and descriptions.
  2. Agent tags are specified as an attribute of the Orchestrator agent. Providing it with all flows tagged as sub-agents can confuse and overwhelm the agent when selecting the appropriate sub-agent.

Sub-agent Card

Input

When invoking a sub-agent, it receives detailed context, including:

  • Its Role and Objectives (as a system message).
  • Domain Data (details on relevant LUs, reference tables, column names, and descriptions) to enable dynamic SQL query construction.
  • A specific List of Tools (other Broadway flows) tailored to the domain.

Logic

A typical sub-agent is composed of logic steps alongside LLMAgent actor usage, providing improved control and reliability.

The LLMAgent actor decides which tools to activate and how to formulate responses based on:

  1. Predefined Domain Data — retrieved using a DBCommand actor
  2. Schema Information — tables relevant for SQL crafting
  3. System Prompt — contains agent goals and behavioral guidelines
  4. User Prompt — the refined user query
  5. Tool List — available tools that the AI can invoke

Use Broadway flow properties to add tags and descriptions to sub-agent flows. The framework uses these properties to select the appropriate agent for each request.

© Copyright 2026 K2VIEW