🐛 Bug Report: Langchain instrumentation adds empty create_agent graph nodes for every middleware hook

Open
#4,628 4 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

@DEVJHAWAR11 is already working on this.

Since Oct 16, 2026.

  • #4629 by @DEVJHAWAR11 — open

Assessment

Difficulty
3/5
Estimated time
1-2 days
Newbie friendliness
58/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
python
Domain
observability

Research direction

Start with _wrap_middleware_hooks in packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/__init__.py and compare it with the middleware hook identity check in LangChain’s factory.py. Reproduce the graph change using the script in the issue, then inspect #4000 for the proposed instrumentation-side fix. Done means instrumentation preserves hook identity, adds no empty graph nodes, and does not increase steps toward recursion_limit.

Written by the indexing model from the issue text.

Description

Which component is this bug for?

Langchain Instrumentation

📜 Description
  • _wrap_middleware_hooks wraps the AgentMiddleware base hooks on the class with wrap_function_wrapper (__init__.py#L298-L322). A wrapt FunctionWrapper returns a new bound wrapper on every attribute access, so AgentMiddleware.before_model is AgentMiddleware.before_model becomes False.
  • create_agent uses that identity to decide which middleware overrides a hook (factory.py#L1166-L1185). After instrumenting, every middleware looks like it overrides all four hooks, so the graph gains an empty before_model / after_model node per middleware per model turn. Each node counts as a step against recursion_limit.
  • In our agents with four wrap_model_call middlewares, a model turn went from about 2 steps to about 10, and runs hit GraphRecursionError around 5x earlier than without tracing.
  • Why we're raising this here rather than with LangChain:
    • there might be an improvement to LangChain's check that could harden their logic but I think the tracing is meant to observe the agent without changing how it runs, and #4000 suggests it can be fixed entirely inside the instrumentation.
    • openllmetry's span wrapper also seems to fire only when the base (no-op) hook runs. A middleware that really overrides before_model replaces the base method, so the wrapper never runs and I don't think real hook overrides get an execute_task span today. Could you confirm whether that matches your understanding?
👟 Reproduction steps
  1. Save the script below as repro.py.
  2. Run it in a clean environment:
    uv run --no-project --with "opentelemetry-instrumentation-langchain==0.62.4" --with "langchain>=1.3" --with opentelemetry-sdk python repro.py
    
from importlib.metadata import version
import sys

from langchain.agents import create_agent
from langchain.agents.middleware import AgentMiddleware, wrap_model_call
from langchain_core.language_models.fake_chat_models import FakeMessagesListChatModel
from langchain_core.messages import AIMessage


@wrap_model_call
def passthrough(request, handler):
    return handler(request)


def build_nodes():
    model = FakeMessagesListChatModel(responses=[AIMessage(content="done")])
    agent = create_agent(model, tools=[], middleware=[passthrough])
    return sorted(agent.get_graph().nodes)


print("python", sys.version.split()[0])
for package in ("langchain", "langgraph", "opentelemetry-instrumentation-langchain"):
    print(package, version(package))

print("before instrument:", build_nodes())
print("identity before:", AgentMiddleware.before_model is AgentMiddleware.before_model)

from opentelemetry.instrumentation.langchain import LangchainInstrumentor

LangchainInstrumentor().instrument()

print("after instrument: ", build_nodes())
print("identity after:", AgentMiddleware.before_model is AgentMiddleware.before_model)
👍 Expected behavior

Instrumenting doesn't change the agent graph. A middleware that only implements wrap_model_call adds no nodes:

after instrument:  ['__end__', '__start__', 'model']
identity after: True
👎 Actual Behavior with Screenshots
python 3.14.0
langchain 1.4.6
langgraph 1.2.15
opentelemetry-instrumentation-langchain 0.62.4
before instrument: ['__end__', '__start__', 'model']
identity before: True
after instrument:  ['__end__', '__start__', 'model', 'passthrough.after_agent', 'passthrough.after_model', 'passthrough.before_agent', 'passthrough.before_model']
identity after: False
🤖 Python Version

3.14.0. Also seen on 3.12.

📃 Provide any additional context for the Bug.
  • Versions where we see it: instrumentation 0.62.1, 0.62.3 and 0.62.4, with langchain 1.2.13, 1.3.15 and 1.4.6.
  • #4000 looks like it addresses exactly this. Running the script above against its head commit (43d14f5, with wrapt<2) keeps the graph unchanged and identity True. The branch is based on 0.59.1, so it may need a rebase onto main for the wrapt 2 changes from #4082.
  • In the meantime we're working around it downstream by calling opentelemetry.instrumentation.utils.unwrap on the eight AgentMiddleware hooks right after instrument().
👀 Have you spent some time to check if this bug has been raised before?
  • I checked and didn't find similar issue
Dominant language
Python
Stars
7.5k
Forks
1.2k
Avg merge
2d 6h
Merged PRs (30d)
2

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from traceloop/openllmetry

All issues in traceloop/openllmetry

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.