diff --git a/app/ai/agent_workstream.py b/app/ai/agent_workstream.py
new file mode 100644
index 0000000..af9a943
--- /dev/null
+++ b/app/ai/agent_workstream.py
@@ -0,0 +1,201 @@
+"""Agent → Workstream integration.
+
+Posts agent messages to the central communication system (kommunikation plugin).
+Supports text, action_card, entity_card, task_card, approval, and miniapp block types.
+All agent messages are marked as AI-generated via transparency metadata.
+"""
+
+from __future__ import annotations
+
+import logging
+import uuid
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.ai.transparency import mark_as_ai_generated
+
+logger = logging.getLogger(__name__)
+
+
+async def post_agent_message(
+    db: AsyncSession,
+    tenant_id: uuid.UUID,
+    agent_id: uuid.UUID,
+    agent_run_id: uuid.UUID,
+    content: str,
+    block_type: str = "text",
+    block_data: dict[str, Any] | None = None,
+    conversation_id: uuid.UUID | None = None,
+) -> uuid.UUID:
+    """Post a message from an agent to the communication system.
+
+    Creates a CommMessage in the agent's conversation channel.
+    If no conversation_id is provided, uses the agent's default channel.
+
+    Args:
+        db: Database session.
+        tenant_id: Tenant ID.
+        agent_id: Agent definition ID.
+        agent_run_id: Agent run ID for traceability.
+        content: Message text content.
+        block_type: Block type (text, action_card, entity_card, task_card, approval, miniapp).
+        block_data: Additional block data (e.g. action buttons, entity reference).
+        conversation_id: Optional conversation to post to. If None, uses agent channel.
+
+    Returns:
+        Message ID.
+    """
+    from app.plugins.builtins.kommunikation.models import CommMessage, CommMessageBlock
+    from app.plugins.builtins.kommunikation.services import send_message
+
+    # Mark as AI-generated
+    ai_metadata = mark_as_ai_generated(content, {
+        "agent_id": str(agent_id),
+        "agent_run_id": str(agent_run_id),
+    })
+
+    # Build block if not plain text
+    blocks: list[dict[str, Any]] = []
+    if block_type != "text" and block_data:
+        blocks.append({
+            "type": block_type,
+            "data": block_data,
+        })
+
+    # Post via kommunikation service
+    message_id = await send_message(
+        db=db,
+        tenant_id=tenant_id,
+        sender_id=agent_id,
+        sender_type="agent",
+        conversation_id=conversation_id or await _get_or_create_agent_channel(db, tenant_id, agent_id),
+        content=content,
+        blocks=blocks,
+        metadata=ai_metadata,
+    )
+
+    logger.info(
+        "Agent %s posted message %s (block_type=%s, run=%s)",
+        agent_id, message_id, block_type, agent_run_id,
+    )
+    return message_id
+
+
+async def _get_or_create_agent_channel(
+    db: AsyncSession, tenant_id: uuid.UUID, agent_id: uuid.UUID
+) -> uuid.UUID:
+    """Get or create a dedicated conversation channel for an agent."""
+    from app.plugins.builtins.kommunikation.models import CommConversation
+
+    # Try to find existing agent channel
+    result = await db.execute(
+        select(CommConversation).where(
+            CommConversation.tenant_id == tenant_id,
+            CommConversation.entity_type == "agent",
+            CommConversation.entity_id == agent_id,
+            CommConversation.deleted_at.is_(None),
+        )
+    )
+    conv = result.scalar_one_or_none()
+    if conv:
+        return conv.id
+
+    # Create new channel
+    conv = CommConversation(
+        tenant_id=tenant_id,       entity_type="agent",
+        entity_id=agent_id,
+        title=f"Agent Channel",
+        conversation_type="channel",
+        is_system=False,
+    )
+    db.add(conv)
+    await db.flush()
+    return conv.id
+
+
+async def post_agent_step(
+    db: AsyncSession,
+    tenant_id: uuid.UUID,
+    agent_id: uuid.UUID,
+    agent_run_id: uuid.UUID,
+    step_number: int,
+    thought: str,
+    action: str | None = None,
+    observation: str | None = None,
+) -> uuid.UUID | None:
+    """Post a ReAct step as an action_card to the workstream.
+
+    Only posts if the agent's trace_mode is 'extended'.
+    """
+    block_data = {
+        "step_number": step_number,
+        "thought": thought[:500],  # Truncate for display
+        "action": action,
+        "observation": (observation or "")[:500],
+    }
+    return await post_agent_message(
+        db=db,
+        tenant_id=tenant_id,
+        agent_id=agent_id,
+        agent_run_id=agent_run_id,
+        content=f"Step {step_number}: {action or 'Thinking...'}",
+        block_type="action_card",
+        block_data=block_data,
+    )
+
+
+async def post_agent_result(
+    db: AsyncSession,
+    tenant_id: uuid.UUID,
+    agent_id: uuid.UUID,
+    agent_run_id: uuid.UUID,
+    final_content: str,
+    total_cost_usd: float,
+    steps_taken: int,
+    status: str,
+) -> uuid.UUID:
+    """Post the final result of an agent run to the workstream."""
+    block_data = {
+        "status": status,
+        "steps_taken": steps_taken,
+        "total_cost_usd": round(total_cost_usd, 6),
+        "run_id": str(agent_run_id),
+    }
+    return await post_agent_message(
+        db=db,
+        tenant_id=tenant_id,
+        agent_id=agent_id,
+        agent_run_id=agent_run_id,
+        content=final_content,
+        block_type="action_card",
+        block_data=block_data,
+    )
+
+
+async def post_approval_request(
+    db: AsyncSession,
+    tenant_id: uuid.UUID,
+    agent_id: uuid.UUID,
+    agent_run_id: uuid.UUID,
+    approval_id: uuid.UUID,
+    action: str,
+    description: str,
+) -> uuid.UUID:
+    """Post an approval request card to the workstream."""
+    block_data = {
+        "approval_id": str(approval_id),
+        "action": action,
+        "description": description,
+        "status": "pending",
+    }
+    return await post_agent_message(
+        db=db,
+        tenant_id=tenant_id,
+        agent_id=agent_id,
+        agent_run_id=agent_run_id,
+        content=f"Approval required: {action}",
+        block_type="approval",
+        block_data=block_data,
+    )
diff --git a/app/plugins/builtins/automation/prebuilt/__init__.py b/app/plugins/builtins/automation/prebuilt/__init__.py
new file mode 100644
index 0000000..be4a7e7
--- /dev/null
+++ b/app/plugins/builtins/automation/prebuilt/__init__.py
@@ -0,0 +1 @@
+"""Pre-built agent definitions for common CRM use cases."""
diff --git a/app/plugins/builtins/automation/prebuilt/contact_enrichment_agent.py b/app/plugins/builtins/automation/prebuilt/contact_enrichment_agent.py
new file mode 100644
index 0000000..4127b34
--- /dev/null
+++ b/app/plugins/builtins/automation/prebuilt/contact_enrichment_agent.py
@@ -0,0 +1,52 @@
+"""Pre-built Contact-Enrichment-Agent.
+
+Enriches contact data by searching for related information.
+"""
+from __future__ import annotations
+import uuid
+from app.plugins.builtins.automation.models import AgentDefinition
+
+CONTACT_ENRICHMENT_SYSTEM_PROMPT = """You are a Contact Enrichment Agent for a CRM system.
+
+Your task is to enrich contact profiles with additional information.
+
+For each contact, you should:
+1. Search for related entities (companies, other contacts)
+2. Check audit history for recent interactions
+3. Find semantic matches in the database
+4. Suggest missing fields that could be filled
+5. Identify potential duplicates
+
+Use the available tools to:
+- Search for related entities (search_related)
+- Get entity history (get_contact_history)
+- Call CRM API for data lookup (call_crm_api)
+
+Output format:
+- Enrichment suggestions as structured data
+- Confidence score for each suggestion
+- Source reference for each piece of information
+
+Do NOT modify contacts. You are advisory only.
+"""
+
+def create_contact_enrichment_agent(
+    tenant_id: uuid.UUID, user_id: uuid.UUID
+) -> AgentDefinition:
+    return AgentDefinition(
+        tenant_id=tenant_id,
+        name="Contact-Enrichment-Agent",
+        description="Reichert Kontaktdaten mit verwandten Informationen an",
+        system_prompt=CONTACT_ENRICHMENT_SYSTEM_PROMPT,
+        llm_model="openai/gpt-4o-mini",
+        tool_ids=["search_related", "get_contact_history", "call_crm_api"],
+        max_steps=8,
+        max_duration_seconds=90,
+        budget_limit_usd=0.30,
+        mode="reactive",
+        is_active=True,
+        temperature=0.2,
+        max_tokens=1500,
+        trace_mode="standard",
+        created_by=user_id,
+    )
diff --git a/app/plugins/builtins/automation/prebuilt/email_triage_agent.py b/app/plugins/builtins/automation/prebuilt/email_triage_agent.py
new file mode 100644
index 0000000..21539c5
--- /dev/null
+++ b/app/plugins/builtins/automation/prebuilt/email_triage_agent.py
@@ -0,0 +1,51 @@
+"""Pre-built E-Mail-Triage-Agent.
+
+Sorts and prioritizes incoming emails automatically.
+"""
+from __future__ import annotations
+import uuid
+from app.plugins.builtins.automation.models import AgentDefinition
+
+EMAIL_TRIAGE_SYSTEM_PROMPT = """You are an E-Mail Triage Agent for a CRM system.
+
+Your task is to sort and prioritize incoming emails for the user.
+
+For each email, you should:
+1. Classify it as: urgent, important, normal, low_priority, or spam
+2. Extract key information: sender, subject, intent, action items
+3. Suggest a response category: reply_needed, forward, archive, delete
+4. Identify any contacts that should be linked
+
+Use the available tools to:
+- Fetch emails for contacts (get_contact_mails)
+- Summarize email threads (summarize_mail_thread)
+- Call CRM API for contact/company data (call_crm_api)
+
+Output format:
+- Provide a structured summary of each email
+- Include priority level and suggested action
+- Be concise but thorough
+
+Do NOT send emails or make changes. You are advisory only.
+"""
+
+def create_email_triage_agent(
+    tenant_id: uuid.UUID, user_id: uuid.UUID
+) -> AgentDefinition:
+    return AgentDefinition(
+        tenant_id=tenant_id,
+        name="E-Mail-Triage-Agent",
+        description="Sortiert und priorisiert eingehende E-Mails automatisch",
+        system_prompt=EMAIL_TRIAGE_SYSTEM_PROMPT,
+        llm_model="openai/gpt-4o-mini",
+        tool_ids=["get_contact_mails", "summarize_mail_thread", "call_crm_api"],
+        max_steps=10,
+        max_duration_seconds=120,
+        budget_limit_usd=0.50,
+        mode="reactive",
+        is_active=True,
+        temperature=0.3,
+        max_tokens=2000,
+        trace_mode="standard",
+        created_by=user_id,
+    )
diff --git a/app/plugins/builtins/automation/prebuilt/follow_up_agent.py b/app/plugins/builtins/automation/prebuilt/follow_up_agent.py
new file mode 100644
index 0000000..1f1512b
--- /dev/null
+++ b/app/plugins/builtins/automation/prebuilt/follow_up_agent.py
@@ -0,0 +1,52 @@
+"""Pre-built Follow-up-Agent.
+
+Reminds about and creates follow-up tasks for contacts.
+"""
+from __future__ import annotations
+import uuid
+from app.plugins.builtins.automation.models import AgentDefinition
+
+FOLLOW_UP_SYSTEM_PROMPT = """You are a Follow-up Agent for a CRM system.
+
+Your task is to identify and create follow-up tasks for contacts.
+
+For each contact, you should:
+1. Check open tasks and calendar entries
+2. Review recent email communication
+3. Identify contacts that need follow-up (no response, overdue tasks, upcoming deadlines)
+4. Suggest follow-up actions (call, email, meeting, task)
+5. Create follow-up tasks when appropriate
+
+Use the available tools to:
+- Get open tasks (get_open_tasks)
+- Get contact emails (get_contact_mails)
+- Call CRM API for task creation (call_crm_api)
+
+Output format:
+- List of contacts needing follow-up with reason
+- Suggested action and timing for each
+- Priority level (urgent, this_week, this_month)
+
+You may create tasks via call_crm_api. Always include a clear description and due date.
+"""
+
+def create_follow_up_agent(
+    tenant_id: uuid.UUID, user_id: uuid.UUID
+) -> AgentDefinition:
+    return AgentDefinition(
+        tenant_id=tenant_id,
+        name="Follow-up-Agent",
+        description="Erstellt und erinnert an Follow-up-Tasks für Kontakte",
+        system_prompt=FOLLOW_UP_SYSTEM_PROMPT,
+        llm_model="openai/gpt-4o-mini",
+        tool_ids=["get_open_tasks", "get_contact_mails", "call_crm_api"],
+        max_steps=8,
+        max_duration_seconds=90,
+        budget_limit_usd=0.30,
+        mode="proactive",
+        is_active=True,
+        temperature=0.4,
+        max_tokens=1500,
+        trace_mode="standard",
+        created_by=user_id,
+    )
diff --git a/app/plugins/builtins/automation/prebuilt/report_agent.py b/app/plugins/builtins/automation/prebuilt/report_agent.py
new file mode 100644
index 0000000..f60c9dc
--- /dev/null
+++ b/app/plugins/builtins/automation/prebuilt/report_agent.py
@@ -0,0 +1,50 @@
+"""Pre-built Report-Agent.
+
+Generates reports from CRM data using search and API tools.
+"""
+from __future__ import annotations
+import uuid
+from app.plugins.builtins.automation.models import AgentDefinition
+
+REPORT_SYSTEM_PROMPT = """You are a Report Agent for a CRM system.
+
+Your task is to generate reports from CRM data.
+
+You can:
+1. Search for contacts, companies, and activities using hybrid search
+2. Call CRM API for structured data (contacts, companies, tasks, calendar)
+3. Aggregate and summarize data into reports
+4. Format reports as markdown with tables and sections
+
+Report types you can generate:
+- Contact activity summary (interactions, emails, tasks per contact)
+- Sales pipeline overview (contacts by status, recent changes)
+- Task completion report (open vs done, overdue, by assignee)
+- Communication summary (email volume, response times)
+- Custom reports based on user request
+
+Use the available tools to gather data, then format a clear, structured report.
+Include relevant metrics, dates, and entity references.
+Be concise but comprehensive. Use markdown formatting.
+"""
+
+def create_report_agent(
+    tenant_id: uuid.UUID, user_id: uuid.UUID
+) -> AgentDefinition:
+    return AgentDefinition(
+        tenant_id=tenant_id,
+        name="Report-Agent",
+        description="Generiert Berichte aus CRM-Daten",
+        system_prompt=REPORT_SYSTEM_PROMPT,
+        llm_model="openai/gpt-4o-mini",
+        tool_ids=["call_crm_api", "hybrid_search"],
+        max_steps=12,
+        max_duration_seconds=180,
+        budget_limit_usd=0.50,
+        mode="reactive",
+        is_active=True,
+        temperature=0.3,
+        max_tokens=3000,
+        trace_mode="standard",
+        created_by=user_id,
+    )