[bot] LangGraph Platform API client (@langchain/langgraph-sdk) not instrumented — no tracing for cloud agent execution

[braintrust-bot]

Summary

@langchain/langgraph-sdk is a dedicated HTTP client SDK for the LangGraph Platform (LangGraph Cloud / hosted agent service). It is a completely separate npm package from the @langchain/langgraph local framework and from @langchain/core. As of v1.9.25 it is actively maintained with its own release cadence.

This package is not instrumented today. The existing langchain.ts config only hooks into @langchain/core's CallbackManager (the local framework callback system). Calls made through @langchain/langgraph-sdk bypass that system entirely.

Gap Details

Package: @langchain/langgraph-sdk
Current latest: v1.9.25
Source: https://github.com/langchain-ai/langgraphjs/tree/main/libs/sdk
Zero matches for langgraph-sdk anywhere in js/src/ (confirmed via grep).

Key execution surfaces

Method	Description
client.runs.create(threadId, assistantId, options)	Start a background run on a thread
client.runs.stream(threadId, assistantId, options)	Stream a run's output (returns AsyncGenerator)
client.threads.stream(threadId, options)	Stream from a thread checkpoint
client.runs.wait(threadId, runId)	Wait for a run to complete and return final output

These are the primary generative-AI execution calls that users make against a deployed LangGraph Platform agent. None are traced.

Why the existing instrumentation does not cover this

The langchain.ts config targets:

// js/src/auto-instrumentations/configs/langchain.ts
module: { name: "@langchain/core", filePath: "dist/callbacks/manager.js" }
functionQuery: { className: "CallbackManager", methodName: "configure" }

@langchain/langgraph-sdk makes direct HTTP requests (via its own BaseClient) to the LangGraph Platform REST API. It does not instantiate @langchain/core models or invoke CallbackManager. The two packages share only a namespace prefix, not a runtime dependency path.

Braintrust Docs Status

The Braintrust LangGraph integration page covers only the local framework (@langchain/langgraph + @langchain/core callbacks). There is no mention of @langchain/langgraph-sdk or LangGraph Platform/Cloud. Users running agents on LangGraph Platform through the SDK will see no spans in Braintrust today.

Suggested Approach

Instrument BaseClient or the specific streaming/run methods in @langchain/langgraph-sdk:

• Entry file to target: dist/root.cjs or dist/index.cjs (CommonJS) / dist/root.js (ESM) — the file where Client, RunsClient, and ThreadsClient are defined
• Methods: RunsClient.stream, RunsClient.create, RunsClient.wait, ThreadsClient.stream
• Channel pattern: consistent with existing naming, e.g. orchestrion:@langchain/langgraph-sdk:runs.stream

Alternatively, a source-replacement patch (similar to the Mastra approach) could wrap the Client constructor to inject tracing.