Workflow Create¶

The Workflow Create tool allows agents to create entire workflows programmatically from a YAML DSL specification. This is a core capability for self-improving agents -- they can design and deploy new workflows without human intervention.

Component type: workflow_create

How It Works¶

The tool accepts a YAML string defining a workflow and compiles it into persisted database objects (workflow, nodes, edges, configurations) in a single transaction. It supports two modes:

Create mode -- define name, trigger, and steps for a new workflow from scratch.
Fork mode -- specify based_on: <slug> to clone an existing workflow and apply patches (add/remove steps, update prompts, add/remove tools, update config).

Model Resolution¶

The DSL compiler supports several strategies for resolving which LLM model to use for agent steps:

Strategy	DSL Syntax	Description
`inherit`	`model: inherit`	Copy the model from the parent agent that owns this tool
`capability`	`model: "gpt-4"`	Find the first LLM credential matching a substring
`credential_id`	`model: { credential_id: 5 }`	Direct credential ID pass-through
`discover`	`model: discover`	Auto-discover the best model from available credentials

Ports¶

Outputs¶

Port	Type	Description
`result`	`STRING`	JSON with `workflow_id`, `slug`, `node_count`, `edge_count`, or `error`

Output Format¶

{
  "success": true,
  "workflow_id": 15,
  "slug": "weather-pipeline",
  "node_count": 4,
  "edge_count": 3
}

Configuration¶

This tool has no configurable settings. Model resolution is handled through the DSL syntax.

Usage¶

Connect this tool to an agent via the green diamond tool handle:

flowchart LR
    WC[Workflow Create] -.->|tool| Agent
    WD[Workflow Discover] -.->|tool| Agent
    Model[AI Model] -.->|llm| Agent
    Trigger[Chat Trigger] --> Agent

Tool Signature¶

workflow_create(dsl: str, tags: str = "") -> str

Parameters:

Parameter	Type	Default	Description
`dsl`	string	(required)	YAML string defining the workflow
`tags`	string	`""`	Optional comma-separated tags to add to the created workflow

Example¶

Create Mode¶

An agent creating a simple chat workflow:

name: "Customer Support Bot"
trigger: chat
steps:
  - name: support_agent
    type: agent
    prompt: |
      You are a customer support agent. Be helpful and professional.
      Answer questions about our products and services.
    model: inherit
    tools:
      - web_search
      - http

The agent invokes:

Tool call: workflow_create(
    dsl="name: 'Customer Support Bot'\ntrigger: chat\nsteps:\n  - name: support_agent\n    type: agent\n    prompt: 'You are a customer support agent...'\n    model: inherit\n    tools:\n      - web_search",
    tags="support,customer"
)
Result: {"success": true, "workflow_id": 15, "slug": "customer-support-bot", "node_count": 4, "edge_count": 3}

Fork Mode¶

Cloning and modifying an existing workflow:

based_on: "customer-support-bot"
patches:
  - type: update_prompt
    step: support_agent
    prompt: |
      You are a technical support agent specializing in API issues.
  - type: add_tool
    step: support_agent
    tool: run_command

DSL Step Types¶

DSL Type	Component Type	Description
`agent`	`agent`	LLM agent with tools
`code`	`code`	Python/Jinja2 code execution
`http`	`http_request`	HTTP request
`switch`	`switch`	Conditional routing
`loop`	`loop`	Array iteration
`workflow`	`workflow`	Subworkflow invocation
`human`	`human_confirmation`	Human confirmation gate
`transform`	`code`	Data transformation (alias for code)

DSL Trigger Types¶

DSL Trigger	Component Type
`chat`	`trigger_chat`
`telegram`	`trigger_telegram`
`manual`	`trigger_manual`
`none`	`trigger_workflow`

Discover Before Creating

Use Workflow Discover first to check if a similar workflow already exists. If a close match is found (score >= 0.50), fork mode is more efficient than creating from scratch.

Model Inheritance

When using model: inherit, the created workflow's agent steps automatically inherit the LLM model and credential from the parent agent that owns the workflow_create tool. This is the simplest approach and avoids hardcoding credential IDs.