Fix token count mismatch in Agents dashboard for nested LangGraph setups

[hectorhdzg]

Problem

In multi-agent LangGraph setups (e.g., graphs with sub-graphs), the Agents dashboard in Azure Monitor shows significantly lower token counts than what appears in the trace spans. For example, a trace showing 6,816 total tokens would only display ~2.2K in the Token Consumption chart.

Root Cause

_is_agent_run() only checked the direct parent when preventing nested agent detection:

if run.parent_run_id and run.parent_run_id in self._agent_run_ids:

In LangGraph, agent-like sub-graphs are often separated from the top-level agent by intermediate chain nodes:

TopAgent (agent) -> node_step (chain) -> SubGraph (wrongly detected as agent) -> LLM

Since SubGraph's direct parent is node_step (not in _agent_run_ids), it was incorrectly treated as a separate agent. LLM tokens under it aggregated to SubGraph's bucket instead of the top-level agent. When SubGraph ended, its tokens were discarded -- agent runs don't roll up to parent agents.

Fix

Changed the direct-parent check to walk the full ancestor chain using the existing _find_agent_ancestor() method:

if self._find_agent_ancestor(run) is not None:
    return False

This ensures any agent-like chain nested anywhere under an existing agent is treated as internal, and all LLM token counts aggregate to the single top-level invoke_agent wrapper span.

[github-actions]

Performance comparison

Threshold: regressions >15.0% on gating scenarios fail the build. Higher ops/s is better; positive Δ means the PR is slower.

Scenario	Gating	Baseline (ops/s)	Candidate (ops/s)	Δ %	Status
azure_monitor_log	yes	45,531.1	45,324.8	+0.46%	✅
azure_monitor_span	yes	209,095.7	209,753.5	-0.31%	✅
otel_log	no	55,132.9	55,163.3	-0.06%	✅
otel_span	no	56,652.4	56,315.8	+0.60%	✅

[Copilot]

Pull request overview

This PR fixes incorrect token aggregation in the Azure Monitor Agents dashboard for multi-agent / nested LangGraph setups by ensuring agent detection treats any agent-like chain under an existing agent (even when separated by intermediate non-agent chain nodes) as internal rather than a new agent.

Changes:

• Update _is_agent_run() to detect nested agents by walking the full ancestor chain via _find_agent_ancestor() instead of only checking the direct parent.
• Add a regression unit test covering a deeply nested “agent-like” sub-graph separated from the top-level agent by an intermediate non-agent chain.
• Minor test formatting cleanup in an existing message/tool-call test case.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/microsoft/opentelemetry/_genai/_langchain/_tracer.py	Prevents incorrectly classifying nested agent-like sub-graphs as separate agents by checking all ancestors for an existing agent.
tests/langchain/test_tracer.py	Adds a regression test ensuring deeply nested agent-like chains under an existing agent are not treated as new agents.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.