diff --git a/src/llama_stack/providers/inline/agents/meta_reference/responses/streaming.py b/src/llama_stack/providers/inline/agents/meta_reference/responses/streaming.py
index 95c690147..a73b6bf68 100644
--- a/src/llama_stack/providers/inline/agents/meta_reference/responses/streaming.py
+++ b/src/llama_stack/providers/inline/agents/meta_reference/responses/streaming.py
@@ -8,7 +8,8 @@ import uuid
 from collections.abc import AsyncIterator
 from typing import Any
 
-from llama_stack.core.telemetry import tracing
+from opentelemetry import trace
+
 from llama_stack.log import get_logger
 from llama_stack.providers.utils.inference.prompt_adapter import interleaved_content_as_str
 from llama_stack_api import (
@@ -77,6 +78,7 @@ from .utils import (
 )
 
 logger = get_logger(name=__name__, category="agents::meta_reference")
+tracer = trace.get_tracer(__name__)
 
 
 def convert_tooldef_to_chat_tool(tool_def):
@@ -1094,8 +1096,10 @@ class StreamingResponseOrchestrator:
                 "server_url": mcp_tool.server_url,
                 "mcp_list_tools_id": list_id,
             }
-            # List MCP tools with authorization from tool config
-            async with tracing.span("list_mcp_tools", attributes):
+
+            # TODO: follow semantic conventions for Open Telemetry tool spans
+            # https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/#execute-tool-span
+            with tracer.start_as_current_span("list_mcp_tools", attributes=attributes):
                 tool_defs = await list_mcp_tools(
                     endpoint=mcp_tool.server_url,
                     headers=mcp_tool.headers,
@@ -1171,9 +1175,9 @@ class StreamingResponseOrchestrator:
         if mcp_server.require_approval == "never":
             return False
         if isinstance(mcp_server, ApprovalFilter):
-            if tool_name in mcp_server.always:
+            if mcp_server.always and tool_name in mcp_server.always:
                 return True
-            if tool_name in mcp_server.never:
+            if mcp_server.never and tool_name in mcp_server.never:
                 return False
         return True
 
diff --git a/src/llama_stack/providers/inline/agents/meta_reference/responses/tool_executor.py b/src/llama_stack/providers/inline/agents/meta_reference/responses/tool_executor.py
index 4f294a979..3822fb5b2 100644
--- a/src/llama_stack/providers/inline/agents/meta_reference/responses/tool_executor.py
+++ b/src/llama_stack/providers/inline/agents/meta_reference/responses/tool_executor.py
@@ -9,8 +9,8 @@ import json
 from collections.abc import AsyncIterator
 from typing import Any
 
-from llama_stack.core.telemetry import tracing
-from llama_stack.log import get_logger
+from opentelemetry import trace
+
 from llama_stack_api import (
     ImageContentItem,
     OpenAIChatCompletionContentPartImageParam,
@@ -39,9 +39,12 @@ from llama_stack_api import (
     VectorIO,
 )
 
+from llama_stack.log import get_logger
+
 from .types import ChatCompletionContext, ToolExecutionResult
 
 logger = get_logger(name=__name__, category="agents::meta_reference")
+tracer = trace.get_tracer(__name__)
 
 
 class ToolExecutor:
@@ -296,8 +299,9 @@ class ToolExecutor:
                     "server_url": mcp_tool.server_url,
                     "tool_name": function_name,
                 }
-                # Invoke MCP tool with authorization from tool config
-                async with tracing.span("invoke_mcp_tool", attributes):
+                # TODO: follow semantic conventions for Open Telemetry tool spans
+                # https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/#execute-tool-span
+                with tracer.start_as_current_span("invoke_mcp_tool", attributes=attributes):
                     result = await invoke_mcp_tool(
                         endpoint=mcp_tool.server_url,
                         tool_name=function_name,
@@ -318,7 +322,7 @@ class ToolExecutor:
                     # Use vector_stores.search API instead of knowledge_search tool
                     # to support filters and ranking_options
                     query = tool_kwargs.get("query", "")
-                    async with tracing.span("knowledge_search", {}):
+                    with tracer.start_as_current_span("knowledge_search"):
                         result = await self._execute_knowledge_search_via_vector_store(
                             query=query,
                             response_file_search_tool=response_file_search_tool,
@@ -327,7 +331,9 @@ class ToolExecutor:
                 attributes = {
                     "tool_name": function_name,
                 }
-                async with tracing.span("invoke_tool", attributes):
+                # TODO: follow semantic conventions for Open Telemetry tool spans
+                # https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/#execute-tool-span
+                with tracer.start_as_current_span("invoke_tool", attributes=attributes):
                     result = await self.tool_runtime_api.invoke_tool(
                         tool_name=function_name,
                         kwargs=tool_kwargs,