Workflow Builder

The visual workflow builder enables teams to design multi-agent pipelines using a drag-and-drop canvas. Built on React Flow (@xyflow/react), it provides a no-code interface for creating, editing, and deploying agent workflows while maintaining full programmatic control through JSON workflow definitions.

Canvas Interface

The workflow builder presents a zoomable, pannable canvas where agents are represented as nodes and data flow is represented as edges between them. The canvas supports:

• Drag-and-drop agent nodes from the sidebar palette
• Edge drawing by dragging from output handles to input handles
• Multi-select for batch operations on nodes
• Minimap for navigating large workflows
• Auto-layout to organize complex pipelines

Node Types

Agent Nodes

Agent nodes represent individual AI agents in the pipeline. Each node displays:

• Agent name and version
• Model provider and model name
• Average latency and cost per execution
• Current status indicator (healthy, degraded, offline)

Double-clicking a node opens its configuration panel where you can adjust the model, prompt version, timeout, and retry settings.

Conditional Edges

Conditional edges route data between agents based on runtime conditions. A conditional edge evaluates an expression against the output of the source agent and routes to different downstream agents based on the result.

{
  "condition": "confidence_score > 0.8",
  "true_target": "detailed_analysis_agent",
  "false_target": "fallback_agent"
}

Parallel Branches

Agents without dependencies on each other execute in parallel automatically. The canvas visually represents parallel branches as side-by-side columns. You can explicitly create parallel branches by connecting a single source node to multiple downstream nodes.

JSON Workflow Definitions

Every workflow created in the visual builder has a corresponding JSON definition that can be exported, version-controlled, and imported programmatically.

{
  "workflow_id": "data-analysis-v3",
  "name": "Data Analysis Pipeline",
  "nodes": [
    {
      "id": "vision",
      "agent": "vision_analysis",
      "version": "v3",
      "model": "gemini-2.5-pro",
      "config": {
        "timeout_ms": 30000,
        "retry_count": 2
      }
    },
    {
      "id": "style",
      "agent": "style_analysis",
      "version": "v3",
      "model": "gpt-5.1",
      "depends_on": ["vision"]
    },
    {
      "id": "critic",
      "agent": "style_critic",
      "version": "v2",
      "model": "deepseek-v3",
      "depends_on": ["style"]
    }
  ],
  "edges": [
    { "from": "vision", "to": "style" },
    { "from": "style", "to": "critic" }
  ]
}

JSON definitions can be used to:

• Deploy workflows programmatically via the API
• Store workflow versions in source control
• Share workflow templates across organizations
• Migrate workflows between environments

Workflow Execution

When a workflow is triggered, the builder shows real-time execution progress via Server-Sent Events (SSE). Each node updates its visual state as it transitions through execution phases:

1. Queued — waiting for upstream dependencies
2. Running — agent is actively processing
3. Streaming — tokens are being generated (for LLM agents)
4. Complete — agent has finished successfully
5. Failed — agent encountered an error (with retry indicator)

The SSE stream delivers AgentEvent payloads that the frontend processes to update the canvas in real time.

const eventSource = new EventSource(
  `/api/v1/workflows/${workflowId}/execute?input=${encodedInput}`
);
 
eventSource.onmessage = (event) => {
  const agentEvent = JSON.parse(event.data);
  updateNodeStatus(agentEvent.agent, agentEvent.type);
};

Template Library

The workflow builder includes a template library with pre-built patterns for common use cases:

Templates can be customized after instantiation and saved as new templates for your organization.

API Endpoints

List Workflows

GET /api/v1/workflows

Returns all workflows for the authenticated organization.

Create Workflow

POST /api/v1/workflows
Content-Type: application/json
 
{
  "name": "My Pipeline",
  "nodes": [...],
  "edges": [...]
}

Creates a new workflow from a JSON definition.

Execute Workflow

POST /api/v1/workflows/{workflow_id}/execute
Content-Type: application/json
 
{
  "input": { ... }
}

Triggers workflow execution and returns an SSE stream of AgentEvent payloads.

All endpoints require a valid JWT and X-Org-ID header.