workflow-creator

name	workflow-creator
description	Create and edit Microsoft Agent Framework workflows from natural-language descriptions. Use when the user wants to build a new workflow, add/remove edges or executors, or modify an existing workflow graph. Outputs a validated workflow definition (YAML + Python code).
license	MIT
compatibility	Requires python3, agent-framework SDK
metadata	{"author":"poc-team","version":"1.0"}
allowed-tools	load_skill read_skill_resource run_skill_script

Workflow Creator Skill

Purpose

This skill enables an orchestrator agent to create, edit, and validate Microsoft Agent Framework workflow definitions from:

1. Natural-language prompts – "Build a workflow that takes customer inquiries, classifies them, and routes to the appropriate agent."
2. Manual editing – Add / remove executors and edges to an existing workflow.

Core Concepts Reference

• Executor: A processing unit (AI agent or custom logic) that receives typed messages, processes them, and produces output. Defined with @handler decorator.
• Edge: Connection between executors. Types: Direct, Conditional, Switch-Case, Fan-out, Fan-in.
• WorkflowBuilder: Constructs the directed graph from executors + edges.

Read references/WORKFLOW_PATTERNS.md for detailed examples.

Workflow

Step 1 – Gather Requirements

Collect the following:

Field	Required	Description
name	yes	Unique snake_case workflow name
description	yes	What the workflow does
executors	yes	List of executor definitions (name, type, handler logic)
edges	yes	List of edges (source → target, optional condition)
start	yes	Name of the starting executor

Step 2 – Generate Workflow Definition

Produce two artefacts:

1. workflow.yaml – declarative definition:

name: <name>
description: <description>
start: <start_executor_name>
executors:
  - name: classifier
    type: agent  # or "function"
    instructions: "Classify customer inquiry into: billing, technical, general"
    model: gpt-4o
  - name: billing_agent
    type: agent
    instructions: "Handle billing inquiries"
    model: gpt-4o
  - name: tech_agent
    type: agent
    instructions: "Handle technical support"
    model: gpt-4o
edges:
  - source: classifier
    target: billing_agent
    condition: "result.category == 'billing'"
  - source: classifier
    target: tech_agent
    condition: "result.category == 'technical'"

2. workflow.py – executable Python code:

from agent_framework import (
    Executor, WorkflowBuilder, WorkflowContext, handler, executor
)

class Classifier(Executor):
    @handler
    async def classify(self, message: str, ctx: WorkflowContext[str]) -> None:
        # Classification logic
        await ctx.send_message(result)

# ... more executors ...

builder = WorkflowBuilder(start_executor=classifier)
builder.add_edge(classifier, billing_agent)
builder.add_edge(classifier, tech_agent)
workflow = builder.build()

Step 3 – Validate

Run the validate_workflow script to check:

• YAML is well-formed and all fields present
• Every edge references valid executor names
• Start executor exists
• Graph is connected (all executors reachable from start)
• No duplicate edges
• Python code has no syntax errors

If validation fails, report errors and suggest fixes.

Step 4 – Persist

Save to generated/workflows/<name>/.

Natural Language → Workflow Generation

When the user describes a workflow in plain language:

1. Identify executors – Extract individual processing steps from the description. Each distinct task becomes an executor.
2. Determine executor types – If the step needs LLM reasoning → agent type. If pure logic → function type.
3. Map edges – Determine data flow. Look for keywords like "then", "if", "route to", "fan out", "collect", "aggregate".
4. Identify conditions – Extract routing conditions for conditional edges.
5. Set start executor – The first step in the process.

Edge Type Selection Guide

User language	Edge type
"then", "next", "followed by"	Direct
"if … then … else"	Conditional
"based on", "depending on", "route to"	Switch-Case
"all of", "in parallel", "at the same time"	Fan-out
"combine", "aggregate", "merge"	Fan-in

Editing Existing Workflows

When editing, load the current workflow.yaml, apply changes, and re-validate:

• Add executor: Append to executors list, add edges as needed
• Remove executor: Remove from list, remove all connected edges
• Add edge: Add to edges list
• Remove edge: Remove from edges list
• Modify condition: Update the condition on the specified edge
• Change instructions: Update executor instructions

Always re-validate after editing.