Lecture Notes: Building Complex Agent Workflows with Flowise

Introduction

• Flowise provides two powerful solutions:
  • Multi-agents
  • Sequential agents
• Aim: Compare these features by creating the same project using both methods.

Multi-Agent Example

• Supervisor Node: Manages task delegation among worker nodes (e.g., Software Developer, Code Reviewer).
• Workflow:
  • The supervisor is delegated to manage tasks.
  • Add a Chat OpenAI LLM to the supervisor.
  • Use settings: GPT40 model, Temperature: 0.4.
  • Connect Chat OpenAI node to supervisor node.
  • Supervisor's task: Decides which team member acts next or terminates the process.

Setting Up Multi-Agent Team

• Nodes:
  • Supervisor node is set up with an LLM.
  • Worker nodes added: Software Developer and Code Reviewer.
• Prompts:
  • Software Developer: Build app based on user requirements using technologies like React, JavaScript, Tailwind CSS.
  • Code Reviewer: Ensure code quality and provide feedback.
• Execution:
  • Tasks are assigned, executed, and feedback is shared between nodes.
  • Supervisor manages back-and-forth communication until task completion.

Limitations of Multi-Agents

• Limited control over application behavior.
• Reliance on supervisor for decision-making.

Sequential Agent Example

• Objective: Replicate the multi-agent application using sequential agents.
• Setup:
  • Start with a new flow: Sequential Agent Software Team.
  • Add a starting node and connect it with a Chat OpenAI node.
  • Model: GPT40 with Temperature: 0.2.
  • Add a state node to keep track of application state.

Building Sequential Agent Workflow

• Supervisor Node:
  • Uses LLM to determine the next node and set a state variable.
  • System prompt instructs supervisor on managing worker conversations.
• Node Logic:
  • Set system and human prompts to guide conversation flow.
  • Extract task decision using JSON structured output.

Condition and Execution

• Condition Node:
  • Determines next action based on LLM output stored in state.
  • Conditions set for Software Developer and Code Reviewer.
• Workers Setup:
  • Nodes for Software Developer and Code Reviewer.
  • Each node completes a task and loops back to the supervisor.

Completing the Workflow

• Loop and End State:
  • Set up loops for continuous back and forth between nodes.
  • Use end state to summarize and conclude the process.

Conclusion

• Example shows how to effectively use sequential agents for complex workflows.
• Additional resources recommended for learning more about sequential agents.