Skip to main content
Agent Flow lets you build complex AI pipelines by visually connecting multiple agents, models, tools, and Knowledge Bases. Similar to n8n or Dify, it’s a visual workflow builder — orchestrate multi-agent pipelines via drag and drop.
Agent Flow is only available when developer mode (developer_mode) is enabled and the agent_flow license is granted. Enable developer mode in admin settings and verify the license condition.

​
Flow vs. Single Agent

AspectSingle AgentAgent Flow
ComplexitySimple Q&AMulti-step pipeline
Agent connectionsStandaloneData passed between agents
Conditional branchingNot possibleBranch via Condition / Router nodes
Data transformationNot possibleExtract Field / Format Text transforms
Safety validationPer agentValidate within the flow via Guardrail node
ReusabilityIndividual callsNest pipelines via Subflow
A single agent suffices for simple Q&A or RAG search. Use a Flow when you need to chain or branch across multiple agents.

​
Flow List

In Workspace > Flow, view all created flows.
Flow list
Each flow card shows:
ElementDescription
NameFlow identifier name
DescriptionPurpose of the flow
Active / Inactive badgeFlow active/inactive state
Node countNumber of nodes in the flow
AuthorShown as “By “
ModifiedLast modified time (relative)

​
Creating a Flow

1

Create a new flow

In Workspace > Flow, click the ”+” button (aria-label: “Create Flow”) at the top-right.Create flow
FieldDescriptionExample
Flow IDLowercase letters, digits, hyphen, underscore (2–50 chars)doc-analysis, 3step-review
NameFlow display name”Document Analysis Flow”
DescriptionPurpose of the flow”Summarizes documents and extracts key points”
Flow ID cannot be changed after creation. Must start with a lowercase letter or digit, using only lowercase letters/digits/hyphen/underscore (regex: ^[a-z0-9][a-z0-9_-]*$).
2

Place nodes

From the Nodes panel on the left, drag desired nodes onto the canvas. Every flow must contain a Start node and an End node.Flow canvas
3

Connect nodes

Drag from a node’s output handle (bottom dot) to the next node’s input handle (top dot). Data flows top to bottom.
Handle positions: top = input, bottom = output. Branching nodes like Guardrail and Condition have multiple output handles at the bottom.
4

Configure nodes

Click a node to open the settings panel on the right. For an Agent node, pick the agent to run; for a Model node, set the LLM model and system/user prompts.Node settings panel
5

Validate and save

The top toolbar has Validate and Save as separate buttons.
  • Validate: Checks for missing Start/End nodes, disconnected nodes, cyclic references, etc.
  • Save: Saves the current flow. This is a separate action and doesn’t auto-validate.

​
Build with AI

Click the “AI Assistant” button (purple lightning icon) at the bottom of the flow editor to open the conversational AI builder. Describe the flow you want in natural language, and the AI auto-generates nodes and connections.

​
Conversational Build Flow

The AI builder doesn’t generate immediately — it asks about key decisions first:
1

Describe your intent

Describe the flow, e.g., “Make a flow that classifies customer inquiries by department”.
2

Answer the AI's questions

The AI asks about guardrail use, routing approach, detail options, etc., as buttons. Click the option you want.Examples:
  • “Use a guardrail?” → [Privacy] [Banned Words] [None]
  • “Which routing approach?” → [Router] [Condition] [Direct]
3

Auto-generation

Once enough info is gathered, the AI generates and places nodes and edges on the canvas.
Asking specifically up front skips the questions. Example: “Build a flow with a PII guardrail → sentiment-analysis router → branch to positive/negative agents”

​
AI Builder Features

  • Conversation history retained: Conversation logs are stored on the flow so you can resume editing later
  • Edit existing flows: Opening AI on an existing flow recognizes the current structure and suggests edits
  • Real-time canvas updates: AI-generated nodes appear on the canvas immediately
  • Resizable panel: Drag the top handle to adjust panel size

​
Node Types

Agent Flow provides 10 node types in 4 categories that you can drag from the Node Palette.
CategoryNodes
GeneralStart, Output
AIAgent, Model
Logic & DataCondition, Router, Merge, Transform
WorkspaceGuardrail, Glossary
See the Node Reference for detailed settings of each node.

​
Running a Flow

​
Run from Chat

  1. Start a new chat
  2. In the model selector, the connection-type filter shows Flow. Pick that.
  3. Select the flow you want and enter a message
Flows appear in the model selector as a connection-type filter, not as a separate tab.

​
Execution Flow

During execution, each node’s status is shown in real time.
StateDescription
RunningNode currently being processed
CompletedNode has finished
SourcesWhen using KBSphere, the retrieved document sources

​
Validation

Items checked by the Validate button in the top toolbar:
CheckTypeDescription
Missing Start nodeErrorAt least 1 Start node required
Missing Output nodeErrorAt least 1 Output node required
Resource not selectedErrorAgent/Model/Guardrail/Subflow nodes have no resource specified
Cyclic referenceErrorCycles between nodes detected (infinite loop risk)
Disconnected nodesWarningNodes not connected to anything

​
AI Auto-build Builder

Build flows by chatting in natural language with the AI chat panel at the bottom of the flow editor. Instead of placing nodes one by one, just describe the workflow you want.

​
How to Use

1

Open the chat panel

In the AI chat area at the bottom of the flow editor, select a model.
2

Describe the workflow in natural language

Type the flow you want.Examples:
  • “Classify customer inquiries, then route technical questions to the tech agent and general questions to the CS agent”
  • “Validate input with a guardrail; if it passes, run Agent A → Agent B in order”
  • “Run 3 agents in parallel and merge the results”
3

Review the AI's response

The AI generates the flow structure and applies it to the editor automatically. The AI may ask follow-up questions; click button choices when shown.
4

Iterate

Continue the conversation to add, change, or delete nodes.
  • “Replace the Router node with a Condition”
  • “Insert Agent C in the middle”
  • “Remove the guardrail node”
The AI builder is only available to users with flow write permission. Conversation history is saved on the flow so you can resume editing.

​
One-shot Generation vs. Conversational Edit

ModeDescriptionWhen to Use
One-shot generationGenerate the entire flow from a one-sentence descriptionBuilding a new flow from scratch
Conversational editAdd/change/delete panels gradually through multi-turn conversationImproving an existing flow
AI-generated flows can be modified manually. It’s effective to use the AI builder to scaffold the structure and then tune detailed settings per node.

​
Export / Import

Export and import are done from the flow editor (FlowEditor) top toolbar.
Click Export in the top toolbar of the flow editor to download as JSON. The export includes flow name, description, node configuration, and connection info.

​
Use Cases

[Start] → [Agent: Document Summary] → [Agent: Key Extraction] → [End]
The summary agent processes user input, then the key extraction agent organizes the main points from the summary.
                     ┌─ True ─→ [Agent: Tech Support] ─┐
[Start] → [Condition] ──────────────────────────────→ [End]
                     └─ False → [Agent: General] ─────┘

State Key: input
Condition Type: Contains
Value: "error"
If the user input contains “error”, route to the tech-support agent; otherwise, route to the general agent.
[Start] → [Agent: DBSphere] → [Transform: Format Text] → [Model: Report Generation] → [End]
The DBSphere agent queries data, the Transform node formats the result with Format Text, and the Model node writes the final report.
                                        ┌─ Pass → [Agent: Q&A] → [End]
[Start] → [Guardrail: PII Check] ─┤
                                        └─ Block → [Transform: Error Message] → [End (Error)]
Block input containing PII; only safe input goes to the agent. When Block Action is set to Continue, the block info can be processed by the Transform node and shown to the user.

​
Access Control

Flows have the same access control as other workspace resources.
OptionDescription
PublicAvailable to all users
PrivateAvailable only to the creator
GroupRead/write access granted to specific groups

​
FAQ

Agent Flow appears in Workspace only when both conditions are met:
  1. Developer mode (developer_mode) is enabled
  2. The agent_flow license is granted
Ask your admin to verify the settings.
All agents registered in the workspace. General mode, KBSphere (enhanced RAG), and DBSphere (database) modes are all supported.
Execution stops at the failing node and an error message is shown. Use an Error Handler node to provide an alternate path on errors. Check node settings (resource selection, parameters) and try again.
Yes — connect agents sequentially, or branch with Condition/Router nodes. Each agent’s output is automatically passed as input to the next.
Starts at the Start node and follows connected nodes in order. Condition nodes branch based on the evaluation result (True or False), Router nodes branch to the matching Route, and Aggregator nodes merge multiple parallel paths.
Use the Export button in the top toolbar to download the JSON, then Import it in the target environment’s editor. Referenced agents, models, and guardrails must exist in the target environment.

​
Next Steps

Build Agents

Configure the agents your flow will use

Configure Guardrails

Set up filters for safer responses inside the flow