feat: Refactor Agent Workbench to Agent Fabric and enhance tool metadata handling

Signed-off-by: Andre Bossard <anbossar@microsoft.com>

Author: abossard

backend/agent_workbench/service.py

@@ -110,13 +110,21 @@ def llm(self) -> Any:
     # Tool introspection
     # ------------------------------------------------------------------
 
-    def list_tools(self) -> list[dict[str, str]]:
+    def list_tools(self) -> list[dict[str, Any]]:
         """Return metadata about all registered tools."""
-        result = []
+        result: list[dict[str, Any]] = []
         for t in self._registry.available_tools():
+            input_schema: dict[str, Any] = {"type": "object", "properties": {}}
+            args_schema = getattr(t, "args_schema", None)
+            if args_schema and hasattr(args_schema, "model_json_schema"):
+                try:
+                    input_schema = args_schema.model_json_schema()
+                except Exception:
+                    input_schema = {"type": "object", "properties": {}}
             result.append({
                 "name": t.name,
                 "description": (t.description or "")[:200],
+                "input_schema": input_schema,
             })
         return result
 

backend/app.py

@@ -30,7 +30,7 @@
 
 # Import unified operation system
 
-# Agent Workbench
+# Agent Fabric
 from agent_workbench import (
     AgentDefinitionCreate,
     AgentDefinitionUpdate,
@@ -191,7 +191,7 @@ async def rest_run_agent():
 
 
 # ============================================================================
-# AGENT WORKBENCH ENDPOINTS
+# AGENT FABRIC ENDPOINTS
 # ============================================================================
 
 _WORKBENCH_UI_OPERATION_NAMES = [
@@ -212,7 +212,7 @@ async def rest_run_agent():
 
 @app.route("/api/workbench/ui-config", methods=["GET"])
 async def workbench_ui_config():
-    """Expose UI-friendly endpoint metadata and enums for Agent Workbench."""
+    """Expose UI-friendly endpoint metadata and enums for Agent Fabric."""
     endpoints: list[dict] = []
     for op_name in _WORKBENCH_UI_OPERATION_NAMES:
         op = get_operation(op_name)
@@ -227,7 +227,7 @@ async def workbench_ui_config():
         })
 
     return jsonify({
-        "module": "agent_workbench",
+        "module": "agent_fabric",
         "version": "1",
         "criteria_types": [criteria.value for criteria in CriteriaType],
         "run_statuses": [status.value for status in RunStatus],

backend/operations.py

@@ -325,18 +325,18 @@ async def op_csv_sla_breach_tickets(
 
 @operation(
     name="workbench_list_tools",
-    description="List tools available for Agent Workbench definitions",
+    description="List tools available for Agent Fabric definitions",
     http_method="GET",
     http_path="/api/workbench/tools",
 )
-async def op_workbench_list_tools() -> list[dict[str, str]]:
+async def op_workbench_list_tools() -> list[dict[str, Any]]:
     """Return all registered tool metadata from the workbench registry."""
     return _get_workbench_service().list_tools()
 
 
 @operation(
     name="workbench_list_agents",
-    description="List all Agent Workbench agent definitions",
+    description="List all Agent Fabric agent definitions",
     http_method="GET",
     http_path="/api/workbench/agents",
 )
@@ -348,7 +348,7 @@ async def op_workbench_list_agents() -> list[dict[str, Any]]:
 
 @operation(
     name="workbench_create_agent",
-    description="Create a new Agent Workbench agent definition",
+    description="Create a new Agent Fabric agent definition",
     http_method="POST",
     http_path="/api/workbench/agents",
 )
@@ -360,7 +360,7 @@ async def op_workbench_create_agent(data: AgentDefinitionCreate) -> dict[str, An
 
 @operation(
     name="workbench_get_agent",
-    description="Get one Agent Workbench agent definition by id",
+    description="Get one Agent Fabric agent definition by id",
     http_method="GET",
     http_path="/api/workbench/agents/{agent_id}",
 )
@@ -372,7 +372,7 @@ async def op_workbench_get_agent(agent_id: str) -> dict[str, Any] | None:
 
 @operation(
     name="workbench_update_agent",
-    description="Update an Agent Workbench agent definition",
+    description="Update an Agent Fabric agent definition",
     http_method="PUT",
     http_path="/api/workbench/agents/{agent_id}",
 )
@@ -387,7 +387,7 @@ async def op_workbench_update_agent(
 
 @operation(
     name="workbench_delete_agent",
-    description="Delete an Agent Workbench agent definition",
+    description="Delete an Agent Fabric agent definition",
     http_method="DELETE",
     http_path="/api/workbench/agents/{agent_id}",
 )
@@ -398,7 +398,7 @@ async def op_workbench_delete_agent(agent_id: str) -> bool:
 
 @operation(
     name="workbench_run_agent",
-    description="Run an Agent Workbench agent with a prompt",
+    description="Run an Agent Fabric agent with a prompt",
     http_method="POST",
     http_path="/api/workbench/agents/{agent_id}/runs",
 )
@@ -410,7 +410,7 @@ async def op_workbench_run_agent(agent_id: str, data: AgentRunCreate) -> dict[st
 
 @operation(
     name="workbench_list_agent_runs",
-    description="List Agent Workbench runs for a specific agent",
+    description="List Agent Fabric runs for a specific agent",
     http_method="GET",
     http_path="/api/workbench/agents/{agent_id}/runs",
 )
@@ -426,7 +426,7 @@ async def op_workbench_list_agent_runs(
 
 @operation(
     name="workbench_list_runs",
-    description="List Agent Workbench runs, optionally filtered by agent id",
+    description="List Agent Fabric runs, optionally filtered by agent id",
     http_method="GET",
     http_path="/api/workbench/runs",
 )
@@ -442,7 +442,7 @@ async def op_workbench_list_runs(
 
 @operation(
     name="workbench_get_run",
-    description="Get one Agent Workbench run by id",
+    description="Get one Agent Fabric run by id",
     http_method="GET",
     http_path="/api/workbench/runs/{run_id}",
 )
@@ -454,7 +454,7 @@ async def op_workbench_get_run(run_id: str) -> dict[str, Any] | None:
 
 @operation(
     name="workbench_evaluate_run",
-    description="Evaluate an Agent Workbench run against its success criteria",
+    description="Evaluate an Agent Fabric run against its success criteria",
     http_method="POST",
     http_path="/api/workbench/runs/{run_id}/evaluate",
 )
@@ -466,7 +466,7 @@ async def op_workbench_evaluate_run(run_id: str) -> dict[str, Any]:
 
 @operation(
     name="workbench_get_evaluation",
-    description="Get evaluation for an Agent Workbench run",
+    description="Get evaluation for an Agent Fabric run",
     http_method="GET",
     http_path="/api/workbench/runs/{run_id}/evaluation",
 )

backend/tests/test_workbench_integration_e2e.py

@@ -89,8 +89,18 @@ async def test_create_run_and_evaluate_agent_with_csv_tool(self) -> None:
                 tools_payload = await tools_resp.get_json()
                 tool_names = [tool["name"] for tool in tools_payload["tools"]]
                 self.assertIn("csv_ticket_stats", tool_names)
+                self.assertIn("csv_ticket_fields", tool_names)
+                self.assertFalse(any(name.startswith("list_task") for name in tool_names))
+                self.assertFalse(any(name.startswith("create_task") for name in tool_names))
                 self.assertFalse(any(name.startswith("workbench_") for name in tool_names))
 
+                list_tickets_tool = next(
+                    item for item in tools_payload["tools"] if item["name"] == "csv_list_tickets"
+                )
+                input_props = list_tickets_tool.get("input_schema", {}).get("properties", {})
+                self.assertIn("status", input_props)
+                self.assertIn("limit", input_props)
+
                 create_payload = {
                     "name": "CSV stats verifier",
                     "description": "E2E check for workbench integration",

backend/usecase_demo.py

@@ -18,9 +18,8 @@
 from typing import Any
 from uuid import uuid4
 
-from pydantic import BaseModel, Field, field_validator
-
 from agents import AgentRequest, agent_service
+from pydantic import BaseModel, Field, field_validator
 
 USECASE_DEMO_AGENT_TIMEOUT_SECONDS = float(
     os.getenv("USECASE_DEMO_AGENT_TIMEOUT_SECONDS", "300")
@@ -134,6 +133,11 @@ def _extract_columns(rows: list[dict[str, Any]]) -> list[str]:
     return columns
 
 
+def _is_sla_breach_prompt(prompt: str) -> bool:
+    normalized = prompt.lower()
+    return "csv_sla_breach_tickets" in normalized or "sla breach" in normalized
+
+
 class UsecaseDemoRunService:
     """In-memory run orchestration with polling-friendly status updates."""
 
@@ -188,17 +192,28 @@ async def _execute_run(self, run_id: str) -> None:
             error=None,
         )
 
-        # Enforce a predictable output block for table rendering.
-        structured_prompt = (
-            f"{run.prompt}\n\n"
-            "Antwortformat:\n"
-            "- Führe die Anfrage mit möglichst wenigen Tool-Aufrufen aus.\n"
-            "- Nutze kompakte fields und sinnvolle limits.\n"
-            "- Fordere notes/resolution nur bei explizitem Bedarf an.\n"
-            "- Gib einen JSON-Codeblock mit {\"rows\": [...]} zurück.\n"
-            "- Falls keine sinnvollen Zeilen existieren, gib {\"rows\": []} zurück.\n"
-            "- Optional danach: kurze Zusammenfassung in 2-4 Stichpunkten."
-        )
+        if _is_sla_breach_prompt(run.prompt):
+            structured_prompt = (
+                f"{run.prompt}\n\n"
+                "Antwortformat für SLA-Breach Usecase:\n"
+                "- Rufe csv_sla_breach_tickets als primäre Quelle auf.\n"
+                "- Bevorzuge einen einzelnen Tool-Aufruf; keine unnötigen Tool-Schleifen.\n"
+                "- Liefere ausschließlich kurze Next-Actions als Markdown (max. 6 Bullet Points).\n"
+                "- Keine JSON-Blöcke zurückgeben.\n"
+                "- Fokus: Priorisierung, Verantwortliche Gruppen, sofortige Eskalationsschritte."
+            )
+        else:
+            # Enforce a predictable output block for table rendering.
+            structured_prompt = (
+                f"{run.prompt}\n\n"
+                "Antwortformat:\n"
+                "- Führe die Anfrage mit möglichst wenigen Tool-Aufrufen aus.\n"
+                "- Nutze kompakte fields und sinnvolle limits.\n"
+                "- Fordere notes/resolution nur bei explizitem Bedarf an.\n"
+                "- Gib einen JSON-Codeblock mit {\"rows\": [...]} zurück.\n"
+                "- Falls keine sinnvollen Zeilen existieren, gib {\"rows\": []} zurück.\n"
+                "- Optional danach: kurze Zusammenfassung in 2-4 Stichpunkten."
+            )
 
         try:
             response = await asyncio.wait_for(

backend/workbench_integration.py

@@ -1,7 +1,7 @@
 """
 Workbench Integration
 
-Wires the project's tools into the Agent Workbench and exposes a
+Wires the project's tools into the Agent Fabric module and exposes a
 singleton WorkbenchService ready to use in app.py.
 
 Separation of concerns:
@@ -28,18 +28,18 @@ def _build_registry() -> ToolRegistry:
 
     Sources:
       1. All @operation-decorated functions via api_decorators.get_langchain_tools()
-         This includes: task management + csv_* ticket operations
+         Exposed to Agent Fabric: csv_* ticket operations only.
 
     The registry is built once at startup and shared with WorkbenchService.
     """
     registry = ToolRegistry()
     try:
         all_tools = get_langchain_tools()
-        user_tools = [
+        ticket_tools = [
             tool for tool in all_tools
-            if not getattr(tool, "name", "").startswith("workbench_")
+            if getattr(tool, "name", "").startswith("csv_")
         ]
-        registry.register_all(user_tools)
+        registry.register_all(ticket_tools)
     except Exception as exc:
         import logging
         logging.getLogger(__name__).warning("Could not load langchain tools: %s", exc)

frontend/src/App.jsx

@@ -81,7 +81,7 @@ export default function App() {
     ...usecaseTabs,
     { value: 'kitchensink', label: 'Kitchen Sink', icon: <DataHistogram24Regular />, path: '/kitchensink', testId: 'tab-kitchensink' },
     { value: 'fields', label: 'Fields', icon: <Info24Regular />, path: '/fields', testId: 'tab-fields' },
-    { value: 'workbench', label: 'Workbench', icon: <Wrench24Regular />, path: '/workbench', testId: 'tab-workbench' },
+    { value: 'workbench', label: 'Agent Fabric', icon: <Wrench24Regular />, path: '/workbench', testId: 'tab-workbench' },
     { value: 'agent', label: 'Agent', icon: <Bot24Regular />, path: '/agent', testId: 'tab-agent' },
   ]
   const activeTab = tabs.find((tab) => location.pathname.startsWith(tab.path))?.value ?? 'csvtickets'

frontend/src/features/usecase-demo/demoDefinitions.js

@@ -17,12 +17,13 @@ Nutze ausschließlich CSV-Daten und nenne die verwendeten Ticket-IDs in Fließte
 
 const SLA_BREACH_DEFAULT_PROMPT = `Call csv_sla_breach_tickets with default parameters (unassigned_only=true, include_ok=false).
 
-Using the returned report, write ONLY a short markdown summary (max 200 words):
-1. State the reference_timestamp used
-2. Group ticket counts by breach_status and assigned_group, and confirm assignee is empty
-3. Recommend actions for the most critical breaches
+Then provide only a concise markdown "next actions" commentary:
+1. Mention the reference_timestamp used.
+2. Identify highest-risk assigned groups (breached first, then at_risk).
+3. Give concrete next actions for the next 30-60 minutes.
+4. Add one short escalation recommendation for unresolved breached tickets.
 
-Do NOT output a JSON block — the frontend fetches and renders the ticket table directly from the API.`;
+Do NOT output a JSON block — the frontend already renders the SLA table from the API.`;
 
 /**
  * Add new demos here to create additional pages without duplicating UI logic.
@@ -116,14 +117,13 @@ export const USECASE_DEMO_DEFINITIONS = [
     defaultPrompt: SLA_BREACH_DEFAULT_PROMPT,
     runHistoryLimit: 25,
     pollIntervalMs: 2000,
-    resultViews: ["sla-breach", "markdown"],
+    resultViews: ["sla-breach", "sla-next-actions"],
     resultSectionTitle: "SLA Breach Results",
     resultSectionDescription:
       "Tickets at risk or already past their SLA threshold, sorted by severity.",
     ticketIdFields: ["ticket_ids", "ticket_id", "ticketIds"],
-    // Tickets are shown inline in the sla-breach result view; disable the separate card.
     matchingTickets: {
-      enabled: true,
+      enabled: false,
       title: "Affected Tickets (Group-Assigned, No Individual Assignee)",
       description:
         'These tickets are routed to a support group but no individual has picked them up. The "Assigned Group" column shows the responsible team; "Assignee" is empty for all.',

frontend/src/features/usecase-demo/resultViews.jsx

@@ -1,21 +1,21 @@
 import {
-  Badge,
-  Button,
-  Checkbox,
-  Spinner,
-  Text,
-  ToolbarButton,
-  Tooltip,
-  makeStyles,
-  tokens
+    Badge,
+    Button,
+    Checkbox,
+    Spinner,
+    Text,
+    ToolbarButton,
+    Tooltip,
+    makeStyles,
+    tokens
 } from '@fluentui/react-components'
 import {
-  ArrowUp24Regular,
-  CheckmarkCircle24Regular,
-  DismissCircle24Regular,
-  Mail24Regular,
-  SelectAllOn24Regular,
-  Warning24Regular,
+    ArrowUp24Regular,
+    CheckmarkCircle24Regular,
+    DismissCircle24Regular,
+    Mail24Regular,
+    SelectAllOn24Regular,
+    Warning24Regular,
 } from '@fluentui/react-icons'
 import { useCallback, useEffect, useMemo, useState } from 'react'
 import ReactMarkdown from 'react-markdown'
@@ -387,6 +387,11 @@ export const RESULT_VIEW_REGISTRY = {
     description: 'Human-readable summary from the run output.',
     render: (props) => <ResultMarkdownView {...props} />,
   },
+  'sla-next-actions': {
+    title: 'Next Actions',
+    description: 'Agent commentary on immediate follow-up actions based on SLA breach data.',
+    render: (props) => <ResultMarkdownView {...props} />,
+  },
   'sla-breach': {
     title: 'SLA Breach Overview',
     description: 'Unassigned tickets color-coded by SLA status. Select tickets to send reminders or escalate.',