Workflow:FlowiseAI Flowise Agentflow V2 Creation
| Knowledge Sources | |
|---|---|
| Domains | AI_Agents, Visual_Programming, LLM_Ops, Workflow_Orchestration |
| Last Updated | 2026-02-12 07:30 GMT |
Overview
End-to-end process for building, executing, and monitoring an AI agent flow using the Flowise V2 agent canvas with support for conditional branching, iteration loops, human-in-the-loop interaction, and AI-powered flow generation.
Description
This workflow covers the creation and execution of Agent Flows in Flowise V2. Unlike simple chatflows (which are linear pipelines), agent flows support advanced control structures including conditional branching, iteration loops, human-input decision points, and multi-agent orchestration. The V2 canvas provides a dedicated node type system with execution status visualization, an AI-powered flow generator that converts natural language descriptions into complete flows, and a hierarchical execution trace viewer for debugging. Agent flows are stored as type AGENTFLOW and can be monitored through the Agent Executions dashboard.
Usage
Execute this workflow when you need to build an autonomous AI agent that goes beyond simple question-answering. This applies when your use case requires conditional logic (routing based on intent or content), iterative processing (looping over items), human approval steps, or multi-agent collaboration. Agent flows are suitable for complex tasks like multi-step research pipelines, customer support routing with escalation, document processing with review gates, and automated report generation with iteration.
Execution Steps
Open the Agentflows page from the sidebar navigation. This page displays all existing agent flows with search, pagination, and card or list view options. The listing distinguishes between V1 (deprecated) and V2 agent flows.
Key considerations:
- Agent flows are scoped to the current workspace
- V1 agent flows are deprecated; new flows use the V2 canvas
- Existing flows can be duplicated or used as starting templates
Step 2: Create New Agent Flow Canvas
Click the "Add New" button to open a blank V2 agent canvas. The system navigates to the dedicated agent flow canvas editor with custom node types (agentFlow, stickyNote, iteration), custom edge rendering with gradient colors, and a specialized node palette containing agent-specific components.
Key considerations:
- The V2 canvas uses distinct node types from the chatflow canvas
- Custom edges display gradient colors between source and target nodes
- Connection validation includes cycle detection to prevent infinite loops
- The canvas supports sticky notes for documentation purposes
Step 3: Add Agent Nodes
Drag agent nodes from the sidebar palette onto the canvas. Available node types include LLM agent nodes, condition nodes (for branching), iteration nodes (for loops), human input nodes (for approval gates), and tool nodes. Alternatively, use the AI-powered flow generator to automatically create a complete flow from a natural language description.
Key considerations:
- Condition nodes support multiple output paths with labeled branches
- Iteration nodes are special container nodes that hold child nodes executed in a loop
- Human input nodes create proceed/reject decision points requiring manual intervention
- The AI generator accepts natural language and produces a complete node graph
- Default templates are available for common patterns (web search, document summarization, multi-agent teams)
Step 4: Configure Node Parameters
Double-click a node or select it to open the edit dialog. Configure the node's input parameters including model selection, prompt templates, tool assignments, and credential references. Parameters are dynamically rendered based on the node type with support for conditional visibility.
Key considerations:
- Each node type exposes different parameter sets
- Credential inputs connect to the centralized credential store
- Parameters support variable references to upstream node outputs
- Some parameters are conditionally displayed based on other selections
- Node versioning alerts users to outdated configurations
Step 5: Connect Nodes with Edges
Drag from output handles to input handles to create connections between nodes. The edge system supports conditional branching (multiple outputs from condition nodes), iteration connections (entry/exit from iteration containers), and human input routing (proceed/reject paths). The system validates connections and prevents cycles.
Key considerations:
- Edges render with gradient colors matching source and target node colors
- Condition edges display numeric branch labels
- Human input edges display proceed/reject labels
- Cycle detection prevents invalid graph structures
- Edges can be deleted via hover-activated delete buttons
Step 6: Save the Agent Flow
Click the save button in the canvas header. The system serializes the complete flow state including all nodes, edges, positions, and iteration container boundaries. The flow is stored with type AGENTFLOW. Conflict detection prevents overwriting concurrent edits.
Key considerations:
- Agent flows are stored as type AGENTFLOW (distinct from CHATFLOW)
- Iteration nodes store parent-child relationships in the serialized state
- Credentials are persisted separately from flow data
- The system tracks dirty state for unsaved change warnings
Step 7: Execute and Monitor
Trigger execution through the chat interface or API. During execution, each node's status is visualized on the canvas with color-coded indicators (in-progress, finished, error, stopped, terminated). After execution, navigate to the Agent Executions page to view a hierarchical tree of the execution trace showing each node's inputs, outputs, timing, token usage, and any errors.
Key considerations:
- Execution status is updated in real time on the canvas via WebSocket or polling
- The execution tree shows parent-child relationships for iteration nodes
- Each node execution record includes: status, start/end time, inputs, outputs, error messages
- Execution details can be shared via a public URL
- Multiple execution runs are tracked and listed chronologically