Agent Flow

Agent Flow lets you build complex AI pipelines by visually connecting agents, models, and knowledge bases. It is a visual workflow builder similar to n8n or Dify — you can compose multi-agent orchestration entirely by drag and drop.

Three conditions must all be met for the Flows tab to appear in your workspace.

The environment must be licensed for agent_flow
Your account must have Flow permission (workspace.agent_flows) — admins always qualify
The address you connect to must be a preview host — Flows is still a preview feature, so the menu appears only on designated development and demo hosts

Because of the third condition, the menu will not appear on a regular production address even with the license and the permission in place. Contact your administrator if you need access.

Flow vs. Single Agent

Aspect	Single Agent	Agent Flow
Complexity	Simple Q&A	Multi-step pipeline
Agent chaining	Runs standalone	Passes data between agents
Conditional branching	Not available	Branch with Condition / Router nodes
Data transformation	Not available	Jinja2 template in the Transform node
Safety checks	Per agent	Guardrail node inside the flow
Human approval	Not available	Mid-flow approval with the Approval Gate node
Batch processing	Not available	Per-item parallel processing with the Iteration node
Reuse	Individual calls	Nest pipelines with Subflow
External invocation	Chat	Chat + webhook endpoint

A single agent is enough for simple Q&A or RAG retrieval. Use a flow when you need to chain several agents sequentially or conditionally.

Flow List

Workspace > Flows shows every flow that has been created.

Each flow card shows:

Element	Description
Active / Inactive badge	Whether the flow is active
Node count	Number of nodes in the flow
Name	Flow identifier name
Description	What the flow is for
Author	Shown as “By {author name}“
Updated	Time of the last change (relative)

Use All / My Flows at the top to narrow the list, or Search Flows to filter by name. The ⋯ menu at the top-right of each card offers Duplicate, Delete, and the Active / Inactive toggle.

An inactive flow is blocked both in the chat model list and for webhook calls. For a flow you want to pause rather than lose, switch it to inactive instead of deleting it — the configuration is preserved.

Creating a Flow

Create a new flow

In Workspace > Flows, click + New Flow at the top right.

Create flow — After entering the create screen

Fill in the basics in the left panel.

Field	Description	Example
Enter flow name	Display name of the flow	”Document Analysis Flow”
Flow ID	2–50 character identifier starting with a lowercase letter or digit	`doc-analysis`
Enter description	What the flow is for	”Summarizes documents and extracts key points”
Stateful	Keep conversation state across runs (on by default)	On

Flow ID availability is checked as you type — a green check when it is free, a red X with “This Flow ID is already in use” when it is taken.

The Flow ID cannot be changed after creation. The on-screen hint mentions lowercase letters, numbers, and hyphens, but underscores (_) are accepted as well (^[a-z0-9][a-z0-9_-]*$).

Start from a template, or from an empty canvas

The empty canvas of a new flow opens with the Start from a template gallery. Picking a card starts you off with nodes and edges already laid out.

Template	What it does
Simple LLM Response	Answers the user input with a single model call
Guarded Agent Answer	Screens the input with a guardrail, then lets an agent answer when it passes
Intent Routing	Routes to differently prompted models depending on question / request / other
Parallel Model Compare	Sends the same input to two models in parallel and merges both answers
Knowledge Base Q&A	Retrieves documents from a knowledge base and answers using that context
Summarize Pipeline	Summarizes with a model, then formats the result with a transform template

Click Build from scratch (✕) at the top right of the gallery to dismiss it and build on an empty canvas.

Templates lay out the skeleton only. Resources such as models, agents, and knowledge bases must be selected on each node before the flow will run.

Picking a template turns the Stateful toggle off, because the template replaces the whole flow configuration. If the flow should keep conversation context, switch Stateful back on after choosing a template.

Place nodes

Drag a node from the Node panel on the left onto the canvas, or click it to drop it at the center of the canvas. As the palette grows, filter it with the Search nodes box at the top.

Flow canvas — Dragged nodes land where you drop them; clicked nodes land at the center of the canvas

Connect nodes

Drag from a node’s output handle (dot on the right) to the next node’s input handle (dot on the left). Data flows left to right.

Handle positions: left = input, right = output. Branching nodes — Condition, Router, Guardrail, Approval Gate — stack multiple output handles vertically on the right. The Start node has no input handle, and the Output node has no output handle.

Configure nodes

Clicking a node opens its settings panel on the right. For an Agent node pick the agent to run; for a Model node pick the LLM and set the system / user prompts. Drag the panel’s left edge to resize it.

Node settings panel 1 — Agent node settings

Node settings panel 2 — Model node settings

See the Node Reference for the settings of each node.

Validate, then save

Validate and Save & Create in the top toolbar are separate actions.

Validate — checks for problems and marks failing nodes on the canvas with a red ring
Save & Create — creates the flow and moves to the edit screen. Saving does not run validation automatically

The dot before the flow name in the toolbar shows validation status.

Dot color	Meaning
Gray	Not validated since last edit
Green	Flow is valid
Amber	Validation errors

Editing the canvas turns the dot gray again — this keeps a stale green result from implying the current graph is still valid.

Save Draft and Publish

Once a flow exists, saving happens in two stages. Saving on the edit screen only updates the draft; publishing is what changes the version that actually runs.

Button	What it does
Save Draft	Saves the current canvas as a draft. The live flow is untouched
Publish	Promotes the draft to live. The previous live version is kept as a version snapshot

When an unpublished draft exists, the Publish button carries an amber Unpublished draft dot and a banner appears above the canvas.

Editing draft — the live flow is unchanged until you publish

Discard draft on the right of the banner asks “Discard this draft and return to the live version?” and then restores the canvas to the live version.

If you only save a draft and leave, nothing changes for your users. Chat and webhook calls keep running the previously published version. Press Publish to actually roll the change out.

This does not apply while creating a flow. The create screen has a single Save & Create button; the split into Save Draft / Publish begins on the edit screen.

Reverting to a previous version

Each publish pushes the previous live version onto a snapshot stack (up to 20, oldest dropped beyond that). Open the Versions tab in the test drawer to review the publish history and press Restore — the snapshot comes back as a draft. Review it on the canvas and press Publish to make the rollback final.

Node Types

Agent Flow provides 13 nodes across 4 categories.

Category	Nodes	Description
General	Start, Output	Flow entry / exit
Agents	Agent, Model, Supervisor	Run an agent, call an LLM directly, coordinate worker agents
Logic & Data	Condition, Router, Merge, Transform, Iteration, Subflow	Branching, routing, combining, transforming, batch processing, flow nesting
Control & Safety	Approval Gate, Guardrail	Wait for human approval, safety filter

See the Node Reference for the settings of each node.

Validation

Use the Validate button in the top toolbar. Failing nodes are marked on the canvas with a red ring.

Errors — must be fixed to run

Check	Description
Missing resource	An Agent / Model / Guardrail / Knowledge / Glossary node with no resource selected
Incomplete supervisor	No supervisor model selected, or zero workers
Incomplete subflow	No flow selected, or the node references itself
Iteration model missing	LLM mode with no model selected
Dangling edge	An edge pointing at a deleted node
Unsupported node type	A type the builder does not handle (imported legacy flows, etc.)
Loop without exit	A cycle with no conditional branch leaving it

Warnings — the flow still runs

Check	Description
No Input node	The first node in the flow is used as the entry point
No Output node	Leaf nodes automatically connect to the flow end
Orphan node	A node connected to nothing (a single-node flow is fine)
Loop detected	The graph contains a cycle

Missing Start / Output nodes are not errors. Implicit entry and exit are supported, so they are reported as warnings only. Adding both is still clearer about intent.

Loops are a supported feature

If a Condition, Router, or Guardrail branch inside the cycle leaves it, the loop is considered safe and only a warning is reported. It is an error only when no branch leaves the cycle at all. Execution is bounded by a superstep limit (50 by default), and loop exits are built by combining {Label.loop_count} with a Greater than condition in a Condition node.

Running a Flow

There are three ways to run a flow.

Test in the editor
Run from Chat
Call by webhook

Test Run in the toolbar opens a test drawer below the canvas. It runs the current canvas as-is, unsaved, so you can check behavior before publishing.The drawer has three tabs.

Tab	Contents
Test Run	Enter test input and press Run. Shows per-node progress and the final output
Run History	Execution history for this flow (owners and admins see all runs; shared users see only their own)
Versions	Publish history and Restore

While running, each node’s status (running / completed / error) appears on the canvas and in the step list at the same time, and clicking a step expands a preview of that node’s output. With a KBSphere agent the number of retrieved sources is shown as well.Drag the top edge of the drawer to resize it.

If the flow contains an Approval Gate, execution stops there and an Approval required card appears. Add a note and press Approve or Reject to resume from that point.

The Webhook endpoint is shown near the bottom of the left panel on the edit screen. Copy URL copies the full address.

POST /api/v1/agent-flows/{flow ID}/invoke

Request field	Description
`input`	Input passed to the flow (required)
`stream`	`true` for an SSE stream; defaults to `false`
`session_id`	Conversation session key. Supplying it keeps state per session

The default response is JSON containing flow_id, session_id, and output.

The endpoint also accepts an API key (Bearer sk-...), which makes it usable from external systems. The caller needs read access to the flow, and calls to an inactive flow are rejected.

Stateful

The Stateful toggle in the left panel of the edit screen controls whether conversation state is kept across runs (on by default for new flows).

Chat runs — state is kept per conversation automatically, regardless of the toggle
Non-chat runs (webhook, embed, scheduled tasks) — the toggle must be on and a session_id supplied for state to carry across calls

Building a Flow with AI

The violet AI Assistant button at the bottom right of the editor opens a conversational builder. Describe the flow you want in plain language and the AI generates the nodes and edges directly onto the canvas.

Pick a model

Choose the model the builder should use in the selector in the panel header.

Describe your intent

Describe the flow you want, e.g. “Build a flow that classifies customer inquiries by department.” Suggestion chips are shown when the panel first opens — click one to start immediately.

Answer the AI's questions

The AI asks follow-ups — whether to use a guardrail, which routing style, and so on — as buttons. Click the option you want.Examples:

“Should we use a guardrail?” → [PII protection] [Banned words] [None]
“Which routing style?” → [Router] [Condition] [Direct connection]

Generate, then refine

Once it has enough information, the nodes and edges are placed on the canvas. Keep the conversation going to refine them.

“Change the router node to a condition”
“Add agent C in the middle”
“Remove the guardrail node”

Being specific up front skips the follow-up questions. For example: “Build a flow with a PII guardrail → sentiment routing → separate agents for positive and negative.”

AI builder features

Conversation history — the transcript is stored with the flow, so you can pick up where you left off
Editing existing flows — opened on an existing flow, it reads the current structure and proposes changes
Live canvas updates — generated nodes appear on the canvas immediately
Resizable panel — drag the handle at the top of the panel

The AI Assistant button is shown only to users with write access to the flow.

Anything the AI generates can still be edited by hand. Sketching the skeleton with AI and tuning each node yourself works well. Results stay a draft until you publish, so there is little risk in experimenting.

Export / Import

Export and import are performed from the flow editor’s top toolbar.

Export
Import

Click Export in the toolbar to download a JSON file. The file contains the flow name, description, node configuration, and connections.

Click Import in the toolbar and upload a previously exported JSON file. Only JSON of the form type: "agent_flow" is accepted.

Agents, models, and guardrails referenced by the imported flow must exist in the current environment. Press Validate first to surface any missing-resource errors.

Use Cases

Example 1: Document analysis flow

[Start] → [Agent: Summarize] → [Agent: Extract key points] → [Output]

A summarization agent processes the user input, then an extraction agent pulls the main points out of the summary.

Example 2: Conditional customer support

                   ┌─ True ──→ [Agent: Tech support] ─┐
[Start] → [Condition] ─────────────────────────────────→ [Output]
                   └─ False → [Agent: General] ───────┘

State Key: input
Condition Type: Contains
Value: "error"

Inputs containing “error” go to the tech support agent; everything else goes to the general inquiry agent.

Example 3: Data report flow

[Start] → [Agent: DBSphere] → [Transform] → [Model: Write report] → [Output]

A DBSphere agent queries the data, the Transform node’s Jinja2 template shapes the result, and a Model node writes the final report.

Example 4: Guardrail flow

                                ┌─ Pass → [Agent: Q&A] → [Output]
[Start] → [Guardrail: PII check] ─┤
                                └─ Block → [Transform: Notice text] → [Output (Error)]

Inputs containing personal information are blocked, and only safe inputs reach the agent. Setting Block Action to Continue lets you shape the block information in a Transform node and explain it to the user.

Example 5: Send-with-approval flow

                                            ┌─ Approve → [Agent: Send] → [Output]
[Start] → [Model: Draft] → [Approval Gate] ─┤
                                            └─ Reject → [Output (Error)]

A model-written draft only reaches the send step after a person reviews it. Set On Reject to Take the reject branch to design a rejection path, or Stop the flow to end there.

Example 6: Per-item parallel processing

[Start] → [Iteration: LLM mode] → [Merge] → [Output]

Given a newline-separated list or a JSON array as input, the Iteration node runs each item through the LLM in parallel and joins the results. Control the load with Concurrency (1–8) and Max Items.

Access Control

Flows use the same access control as other workspace resources. For flows, both the create and the edit screen open the access control modal from the lock-icon [Access] button in the top toolbar — unlike other resources, there is no inline permission area on the screen itself. Pick the sharing scope with the Access permission selector in the modal.

Option	Description
Everyone	Readable by all users
Selected groups only	Available to members of the groups listed under Accessible groups

A group in the table has read access by that fact alone; check the Editable column to grant write access as well.

A newly created flow is readable by everyone by default. Along with channels, this is one of the few exceptions in the workspace — to narrow it, open the flow after creating it and switch Access permission to Selected groups only yourself.

The Everyone option is shown to admins only. Regular users see just Selected groups only.

You must press Apply in the access control modal. Selecting groups and closing the modal does not save the change. After applying, you still need to press Save & Create (create) or Save Draft · Publish (edit) to persist it.

Organization units cannot be selected on this screen. To open access to an organization unit, an administrator must link the organization unit to a group — users in that organization are then treated as members of the group. See Sharing & Permissions for details.

FAQ

I can't see the Flows menu

I saved, but chat still behaves the old way

Save Draft on the edit screen does not change the live flow. You must press Publish for chat and webhook calls to pick it up.If the “Editing draft” banner is above the canvas, or the Publish button has an amber dot, the change is still unpublished.

Which agents can I use in a flow?

Every agent registered in the workspace. Standard mode, KBSphere (enhanced RAG), and DBSphere (database) modes are all supported.

What happens when a flow hits an error?

By default, execution stops at the failing node and an error message is shown.To handle it differently per node, open the Error handling & retry section at the bottom of the settings panel and set On error (Stop the flow / Skip and continue / Use default value) along with Max retries, Retry interval, and Timeout. Every node except Start has this section.

There is no separate Error Handler node.

Can I run the same node repeatedly?

Two ways:

Iteration node — for processing many items at once in parallel (concurrency 1–8)
Loop edges — a cycle that returns to an earlier node based on a condition. Fully supported, provided one Condition / Router / Guardrail branch leaves the loop

Loop execution is bounded by a superstep limit (50 by default), and the exit is built by combining {Label.loop_count} with a Greater than condition.

Can one flow connect several agents?

Yes — chain agents sequentially or branch with Condition / Router nodes. Each agent’s output is passed automatically as the next agent’s input.To have one model split work across several agents and combine the results, use the Supervisor node.

In what order does a flow execute?

Execution starts at the Start node and follows the connected nodes in order. A Condition node branches to the true or false path based on its evaluation, and a Router node branches to the matching route. A Merge node is where parallel paths join.Without a Start node the first node becomes the entry point; without an Output node the leaf nodes become the exit.

How do I roll back to an earlier version?

Open the Versions tab in the test drawer, review the publish history, and press Restore. Restore brings that version back as a draft, so review it on the canvas and press Publish to make the rollback final.Up to 20 snapshots are retained.

How do I move a flow to another environment?

Use Export in the editor toolbar to download a JSON file, then Import it in the target environment’s editor. The referenced agents, models, and guardrails must exist there too — press Validate after importing to confirm.

Next Steps

Node Reference

Settings and State keys for all 13 nodes

Create an Agent

Configure the agents your flow will use

Configure Guardrails

Set up filters for safe responses inside a flow

Sharing & Permissions

How the access control modal works

Workspace

AI Agents

Knowledge & Data

Safety

Customization

Flow vs. Single Agent

Flow List

Creating a Flow

Save Draft and Publish

Reverting to a previous version

Node Types

Validation

Errors — must be fixed to run

Warnings — the flow still runs

Loops are a supported feature

Running a Flow

Stateful

Building a Flow with AI

AI builder features

Export / Import

Use Cases

Access Control

FAQ

Next Steps

Node Reference

Create an Agent

Configure Guardrails

Sharing & Permissions

​Flow vs. Single Agent

​Flow List

​Creating a Flow

​Save Draft and Publish

​Reverting to a previous version

​Node Types

​Validation

​Errors — must be fixed to run

​Warnings — the flow still runs

​Loops are a supported feature

​Running a Flow

​Stateful

​Building a Flow with AI

​AI builder features

​Export / Import

​Use Cases

​Access Control

​FAQ

​Next Steps

Node Reference

Create an Agent

Configure Guardrails

Sharing & Permissions

Flow vs. Single Agent

Flow List

Creating a Flow

Save Draft and Publish

Reverting to a previous version

Node Types

Validation

Errors — must be fixed to run

Warnings — the flow still runs

Loops are a supported feature

Running a Flow

Stateful

Building a Flow with AI

AI builder features

Export / Import

Use Cases

Access Control

FAQ

Next Steps