fix(langchain): emit invoke_agent for create_agent, including nested agents

[AgentGymLeader]

What

Classifies create_agent runs by having the compiled graph announce itself, instead of inferring from callback metadata.

How this got here

The first version of this PR read ls_integration / lc_agent_name off the callback metadata. That does not work: when an agent is invoked from inside another agent's tool with the config forwarded, LangChain merges the caller's config over the inner agent's own, so those fields describe the outer agent. We could not find any callback field that separates the two — serialized is None, langgraph_checkpoint_ns does not deepen at the nested root, and tags / kwargs["name"] do not distinguish an agent from any other named runnable.

@lmolkova reached the same conclusion independently and supplied the mechanism that does work: a compiled graph's own bound config is never shadowed, so patching Pregel.stream / astream lets the graph announce itself on a context stack that the handler claims when the root run starts. That commit is hers, cherry-picked here with authorship preserved.

What changed

create_agent_ancestry, the has_create_agent_marker plumbing, and the ls_integration / lc_agent_name classification branches are gone. agent_context.py holds the announcement stack; claim_agent() takes only the innermost unclaimed entry, so nesting layers without any ancestry walk.

Behaviour change worth noting: a nested agent now gets its own invoke_agent span carrying its own gen_ai.agent.name, falling back to LangGraph when unnamed. The metadata version suppressed it, because the marker it saw was inherited and therefore untrustworthy. With ground truth there is nothing to suppress.

Deprecated create_react_agent and create_tool_calling_executor compile their own graphs and carry no bound-config marker, so they are tagged at the factory instead.

Tests

The classification tests used to construct the callback handler directly against a MagicMock telemetry handler. They asserted the shape of invoke_local_agent calls, which meant they passed for classifications that emitted no usable span at all. They now drive real invocations through instrument() against a real tracer provider and assert on exported spans — names, parent/child nesting, and gen_ai.agent.name. The nesting cases assert the actual parent chain, which is the only form in which they mean anything.

Seven tests that existed only to check the removed inference are deleted.

Verified on both the latest and oldest supported LangChain: 246 passed and 235 passed, 5 skipped respectively. langgraph.graph is imported behind the version guard because it is not importable on the oldest.

Pregel.invoke and ainvoke drive their runs through self.stream / self.astream, so the patch covers non-streaming invocation too.

[opentelemetry-pr-dashboard]

Pull request dashboard status

Waiting on the author · refreshed 2026-08-22 15:51 UTC

Respond to 6 review items (e.g. link a commit, explain why not, ask a follow-up):

• Inline threads: 1, 2, 3, 4, 5
• Top-level threads: 6

Status above doesn't look right?

• Just replied or pushed? Anything around or after the refresh time above may not be picked up yet — give it a few minutes.
• Should this be with reviewers? Comment /dashboard route:reviewers to route it to them.
• Anything wrong — including the routing? Report it with what you expected; it helps us improve the dashboard.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

This PR refines agent span classification for LangChain/LangGraph runs by using LangChain’s lc_agent_name signal while avoiding false positives for internal LangGraph nodes.

Changes:

• Add lc_agent_name-aware agent detection logic in classify_chain_run via _has_agent_signals.
• Avoid misclassifying internal LangGraph nodes as agents by comparing against the callback’s explicit run name.
• Add unit tests covering top-level agents, nested agents, internal nodes, and suppression behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/operation_mapping.py	Updates agent-signal detection to incorporate lc_agent_name with LangGraph-aware heuristics.
instrumentation/opentelemetry-instrumentation-genai-langchain/tests/test_operation_mapping.py	Adds tests validating new classification behavior and suppression rules.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[AgentGymLeader]

@lmolkova Pushed 42aaaf4.

• Conformance: updated the existing scenario. It uses create_agent now and expects invoke_agent, green under weaver 0.25.1.
• ls_integration: agreed, that's the marker. A run is a root when it carries it and no ancestor does, so the default form emits invoke_agent now.
• run_name override: dropped the name comparison. It compares against the nearest marked ancestor now, using the parent runs the handler already tracks.
• resolve_agent_name: checks lc_agent_name before the run name now.

Not fixed: agents nested inside another agent I couldn't do from callback metadata; the measurement and three options are in the PR body.

I resolved your four threads and the two Copilot ones. Reopen anything I closed too early.

[AgentGymLeader]

@lmolkova Folding your prototype in, and reworking the corpus tests through instrument() against exported spans.

Hadn't clocked that the mock made those tests pass with no span at all.

[Copilot]

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.

[lmolkova]

thanks! a few more minor comments. Please update PR description to describe what PR does, it does not need to describe a history of changes it went through

[lmolkova]

A nested agent still gets no span when the outer agent is named via metadata["agent_name"] - the inner run inherits that name, so it is neither equal to declared_agent_name nor different from the parent's, and this branch drops the layer.

outer = create_agent(model, [delegate], name="outer").with_config(
    metadata={"agent_name": "math_agent"}
)
outer.invoke({"messages": [("user", "start")]})
# execute_tool delegate, invoke_agent math_agent - inner agent missing

Suggested change

	suggested_agent_name == declared_agent_name
	declared_agent_name

That restores the layer with all existing tests still passing, but the inner span is then named math_agent. The name precedence in resolve_agent_name needs the same look: inherited metadata outranks the graph's own declared name.

[lmolkova]

This strict xfail asserts nested agents do not work, while test_nested_agent_gets_its_own_span_under_the_calling_tool asserts they do. It only fails because it never starts the instrumentor - add start_instrumentation and it XPASSes. Drop it.

[lmolkova]

The async path has no tests - every new test drives invoke(). wrap_astream is the trickier half (contextvars across __anext__), so it needs at least the root, nested and concurrent-gather cases the sync side has.