Python: execute_tool OTel spans intermittently lost due to missing context propagation in asyncio.gather

[rongrong77]

Description

When an agent uses multiple tools in a single turn, execute_tool spans are intermittently missing from exported traces. The root cause is that _try_execute_function_calls in
_tools.py
executes tool calls concurrently via asyncio.gather without copying the OTel context into each task. This causes span context race conditions — some execute_tool spans fail to attach to their parent and are silently dropped.

The issue is non-deterministic: the same agent invocation sometimes produces complete traces (with all execute_tool spans) and sometimes produces incomplete traces (missing one or more execute_tool spans).

Expected behavior: Every tool execution should produce a correctly-parented execute_tool span, regardless of concurrent execution.

Proposed fix: Wrap each concurrent task with contextvars.copy_context() to ensure OTel context isolation:

Code Sample

from dotenv import load_dotenv
load_dotenv()

import os
os.environ["OTEL_EXPORTER_OTLP_TRACES_ENDPOINT"] = os.getenv("AAMS_OTLP_TRACES_ENDPOINT")
os.environ["OTEL_EXPORTER_OTLP_HEADERS"] = f"x-api-key={os.getenv('AAMS_API_KEY')}"
os.environ["OTEL_EXPORTER_OTLP_PROTOCOL"] = "http/protobuf"

from agent_framework.observability import configure_otel_providers
configure_otel_providers(enable_sensitive_data="true")

from agent_framework import Agent
from agent_framework.openai import OpenAIChatClient

@tool(approval_mode="never_require")
async def tool_a(city: str) -> str:
    return f"Result A for {city}"

@tool(approval_mode="never_require")
async def tool_b(city: str) -> str:
    return f"Result B for {city}"

client = OpenAIChatClient(model="gpt-4o-mini", api_key=os.getenv("API_KEY"), base_url="...")
agent = Agent(client=client, name="TestAgent", instructions="...", tools=[tool_a, tool_b])

# When LLM returns parallel tool calls (tool_a + tool_b),
# sometimes only one execute_tool span appears in the trace.
result = await agent.run("Call both tools", session=agent.create_session())

Error Messages / Stack Traces

No error is raised. The issue is silent span loss. In the exported trace, execute_tool spans are intermittently missing. Example:

Expected spans: invoke_agent → chat → execute_tool tool_a + execute_tool tool_b → chat
Actual (intermittent): invoke_agent → chat → execute_tool tool_a → chat (tool_b span missing)

Root cause location: 
_tools.py
, function _try_execute_function_calls, line:

execution_results = await asyncio.gather(*[
    invoke_with_termination_handling(function_call, seq_idx)
    for seq_idx, function_call in enumerate(function_calls)
])
Python Version: 3.12

Package Versions

agent-framework-core==1.1.0 agent-framework-openai==1.1.0 opentelemetry-sdk==1.40.0 opentelemetry-exporter-otlp-proto-http==1.40.0

Python Version

3.12

Additional Context

No response