diff --git a/docs/INDEX.md b/docs/INDEX.md
index ad15e2e..1861ffe 100644
--- a/docs/INDEX.md
+++ b/docs/INDEX.md
@@ -151,6 +151,93 @@ Handlers (parallel, different entities):
    └─ Run 3: Already synced → Early return (idempotent!)
 ```
 
+---
+
+### Steps vs. Utils: Parallel vs. Sequential
+
+**Separation Principle: "Events for Parallelism, Functions for Composition"**
+
+```
+╔═══════════════════════════════════════════════════════╗
+║  Choose Architecture by Execution Model:              ║
+║                                                       ║
+║  ✅ Separate Steps → When parallel possible          ║
+║     → Communicate via Events                          ║
+║     → Own lock scope, independent retry               ║
+║                                                       ║
+║  ✅ Shared Utils → When sequential required           ║
+║     → Call as function (need return values)           ║
+║     → Reusable across multiple steps                  ║
+╚═══════════════════════════════════════════════════════╝
+```
+
+**Decision Matrix:**
+
+| Requirement | Architecture | Communication | Example |
+|------------|--------------|---------------|---------|
+| **Can run in parallel** | Separate Steps | Events | Document sync + Collection create |
+| **Needs return value** | Utils function | Function call | `get_or_create_collection()` |
+| **Reusable logic** | Utils function | Import + call | `should_sync_to_xai()` |
+| **One-time handler** | Inline in Step | N/A | Event-specific parsing |
+
+**Examples:**
+
+```python
+# ✅ GOOD: Parallel → Separate Steps + Events
+# steps/collection_manager_step.py
+async def handle_entity_update(event_data, ctx):
+    """Creates xAI Collection for entity"""
+    collection_id = await create_collection(entity_type, entity_id)
+    
+    # Fire events for all linked documents (parallel processing!)
+    for doc in linked_docs:
+        await ctx.emit("cdokumente.update", {"entity_id": doc.id})
+
+# steps/document_sync_step.py  
+async def handle_document_update(event_data, ctx):
+    """Syncs document to xAI"""
+    # Runs in parallel with collection_manager_step!
+    await sync_document_to_xai(entity_id)
+
+# ✅ GOOD: Sequential + Reusable → Utils
+# services/document_sync_utils.py
+async def get_required_collections(entity_id):
+    """Reusable function, needs return value"""
+    return await _scan_entity_relations(entity_id)
+
+# ❌ BAD: Sequential logic in Step (not reusable)
+async def handle_document_update(event_data, ctx):
+    # Inline function call - hard to test and reuse
+    collection_id = await create_collection_inline(...)
+    await upload_file(collection_id, ...)  # Needs collection_id from above!
+```
+
+**Guidelines:**
+
+1. **Default to Steps + Events** - Maximize parallelism
+2. **Use Utils when:**
+   - Need immediate return value
+   - Logic reused across 2+ steps
+   - Complex computation (not I/O bound)
+3. **Keep Steps thin** - Mostly orchestration + event emission
+4. **Keep Utils testable** - Pure functions, no event emission
+
+**Code Organization:**
+```
+steps/
+├─ document_sync_event_step.py      # Event handler (thin)
+├─ collection_manager_step.py       # Event handler (thin)
+└─ vmh/
+   └─ entity_link_webhook_step.py   # Webhook handler (thin)
+
+services/
+├─ document_sync_utils.py           # Reusable functions
+├─ xai_service.py                   # API client
+└─ espocrm.py                       # CRM client
+```
+
+---
+
 ### Bidirectional Reference Pattern
 
 **Principle: Every sync maintains references on both sides**