Issue: LangSmith tracing does not capture subagent tool calls in Claude Agent SDK

[Z-Jr10-K]

Issue you'd like to raise.

When using configure_claude_agent_sdk() with Claude Agent SDK, tool invocations made by subagents (defined via agents and mcp_servers) are not recorded in LangSmith traces.
Only the main agent conversation appears.

Reproduction

import asyncio
from typing import Any
from claude_agent_sdk import (
    ClaudeAgentOptions,
    ClaudeSDKClient,
    AgentDefinition,
    tool,
    create_sdk_mcp_server
)
from langsmith.integrations.claude_agent_sdk import configure_claude_agent_sdk

configure_claude_agent_sdk()

@tool("add_numbers", "Add two integers.", {"a": int, "b": int})
async def add_numbers(args: dict[str, Any]) -> dict[str, Any]:
    a, b = args["a"], args["b"]
    return {"content": [{"type": "text", "text": f"{a} + {b} = {a + b}"}]}

math_helper = create_sdk_mcp_server(name="math-helper", version="1.0.0", tools=[add_numbers])

async def main():
    subagent = AgentDefinition(
        description="Performs arithmetic tasks.",
        prompt="You are a math specialist.",
        tools=["add_numbers"],
    )

    options = ClaudeAgentOptions(
        mcp_servers={"math-helper": math_helper},
        allowed_tools=["mcp__math-helper__add_numbers"],
        agents={"math-helper": subagent},
        system_prompt="You are the main orchestrator. Delegate math to math-helper.",
    )

    async with ClaudeSDKClient(options=options) as client:
        await client.query("Ask the math-helper agent to add 3 and 5.")
        async for msg in client.receive_response():
            print(msg)

if __name__ == "__main__":
    asyncio.run(main())

Observed

• Only main claude.conversation trace appears in LangSmith.
• No trace for subagent math-helper or its tool add_numbers.

Example output:

LangSmith should record nested traces for subagents and their tool calls, e.g.:

claude.conversation
 └── math-helper
      └── Tool: add_numbers(a=3, b=5)

Environment

Component	Version
claude-agent-sdk	0.1.4
langsmith	0.4.38
Python	3.13.3
Model	kimi-k2-0905-preview

Suggestion:

No response

[angus-langchain]

Hi @Z-Jr10-K,

I wasn't able to repro this, see trace below. It's important to note that tools are run in a separate async context that destroys contextvars, so we use thread-local context to properly nest the sub-agent call. In order for this to work correctly, the async context variables need to propagate properly through the entire execution chain.

https://smith.langchain.com/public/aa2d9fc8-0fce-4767-ad1b-9e03c2924986/r

Could you share how you're running this code (script, notebook, REPL), and if running directly resolves the issue?

One thing I notice is you're using kimi-k2-0905-preview rather than an Anthropic model; can you confirm if this issue occurs with Claude models as well?

[trevorprater]

Reproducible with Google Vertex AI - Environment Details

@angus-langchain - I can consistently reproduce this issue, and I suspect the key difference is that I'm using Google Vertex AI instead of the direct Anthropic API.

Environment Details

• Claude Agent SDK: 0.1.4
• LangSmith SDK: 0.4.38
• Python: 3.13.2
• Provider: Google Vertex AI (not direct Anthropic API)
  • CLAUDE_CODE_USE_VERTEX=1
  • Enterprise GCP project
• Model: claude-sonnet-4-5@20250929
• MCP Server Type: HTTP (remote server)
• Execution: Standalone Python script (also tested with FastAPI/uvicorn - same issue)
• Environment: Corporate network with proxy/SSL inspection

Observable Symptoms

What Works ✅

• LangSmith traces are created and uploaded successfully
• Main conversation flow is captured
• HTTP POST requests to /api/v1/runs/multipart return 202 Accepted
• Multiple nested trace IDs are created
• Conversation latency tracked correctly

What Doesn't Work ❌

• MCP tool calls are completely invisible in the trace tree
• Tool inputs and outputs are not captured
• Token usage shows 0 tokens / $0.00 despite LLM calls executing
• No nested spans for tool execution appear

Reproducible Example

Test case:

from claude_agent_sdk import ClaudeSDKClient, ClaudeAgentOptions
from langsmith.integrations.claude_agent_sdk import configure_claude_agent_sdk

# Configure LangSmith
configure_claude_agent_sdk()

# Create client with HTTP MCP server
options = ClaudeAgentOptions(
    mcp_servers={
        "atlas": {
            "type": "http",
            "url": "https://internal-mcp-server.example.com"
        }
    },
    system_prompt="Use the Atlas MCP tools to create alerts.",
    permission_mode="bypassPermissions"
)

async with ClaudeSDKClient(options=options) as client:
    await client.query("Create a test alert with ID 'TEST-001'")
    async for msg in client.receive_response():
        print(msg)

Observed behavior:

• Claude's final output: "Perfect! I've successfully created a test alert with ID 'TEST-001'..."
• LangSmith trace shows: Main conversation + 3 LLM calls, but zero tool calls
• The mcp__atlas__create_alert tool that Claude clearly executed is completely missing from the trace

Hypothesis: Vertex AI Async Context Issue

I believe Vertex AI's extra API layer breaks async context propagation:

Direct API path (your working setup?):

ClaudeSDKClient → api.anthropic.com → response

Vertex AI path (my broken setup):

ClaudeSDKClient → Vertex AI SDK → GCP API → Anthropic backend → response

The Vertex AI SDK likely creates new async contexts during authentication/proxying that destroy the contextvars LangSmith uses for trace tracking.

Questions for Debugging

1. What API provider did you test with? (Direct Anthropic API, Bedrock, or Vertex AI?)
2. Can you try reproducing with Vertex AI? This would confirm if it's provider-specific
3. What MCP server types did you test? (stdio, HTTP, SSE?) - I'm using HTTP
4. Python version? You mentioned checking execution environment - I'm on 3.13.2

Diagnostic Logs

LangSmith trace creation is working:

urllib3.connectionpool - DEBUG - POST /api/v1/runs/multipart HTTP/1.1" 202 34
langsmith.client - DEBUG - Sending compressed multipart request
langsmith.client - DEBUG - Tracing control thread func compress parallel called

Multiple trace IDs are created (suggesting nested contexts), but tool calls never appear in the uploaded traces.

LangSmith observability is critical for us to debug and improve our agent's decision-making, but without tool call visibility, we're essentially flying blind.

Workaround?

Is there a recommended way to manually propagate context for Vertex AI environments? Or any guidance on getting this working with GCP's Vertex AI?

Thanks for investigating this!

[Z-Jr10-K]

@angus-langchain - Thanks for the clarification! I can confirm this is provider-specific.

I’m running the Claude Agent SDK with Kimi’s kimi-k2-0905-preview model, which (like Vertex AI) proxies requests and seems to break async context propagation.

Everything works correctly with Anthropic models, but tool traces disappear under Kimi environments.
It’d be great to document this behavior or provide a recommended workaround.

Appreciate the help and confirmation!