LoongSuite DeepAgents Instrumentation

LoongSuite instrumentation for DeepAgents.

Installation

pip install loongsuite-instrumentation-deepagents

Usage

from opentelemetry.instrumentation.deepagents import DeepAgentsInstrumentor

DeepAgentsInstrumentor().instrument()

from deepagents import create_deep_agent

When instrumenting manually, call DeepAgentsInstrumentor().instrument() before importing or binding deepagents.create_deep_agent. Auto-instrumentation runs before application imports and does not need this extra ordering step.

What it does

This instrumentation patches deepagents.graph.create_deep_agent so the final graph returned by DeepAgents is marked with _loongsuite_react_agent = True. It also marks the graph as a DeepAgents agent and injects the same flags into call-time RunnableConfig metadata for invoke, ainvoke, stream, and astream.

The marker is consumed by loongsuite-instrumentation-langchain, which routes the DeepAgents root graph span as an AGENT span instead of a generic CHAIN span.

The LangChain tracer also maps each DeepAgents model decision node to a react step span. Multi-round tool flows end the previous step with gen_ai.react.finish_reason = "tool_calls" and the final step with "stop". This package intentionally does not create separate ENTRY spans or GenAI metrics.

For DeepAgents skills, the framework first loads skill metadata through SkillsMiddleware.before_agent and exposes the available skills in the system prompt. Loading the full skill instructions happens when the agent calls the built-in filesystem tool to read the skill file, for example read_file(file_path="/skills/foo/SKILL.md"). That execute_tool read_file span is annotated with gen_ai.skill.name, gen_ai.skill.id, gen_ai.skill.description, and gen_ai.skill.version when the file path matches a registered top-level SKILL.md. Reading helper files under the same skill directory is recorded as a normal file read, not as a skill load.

Compatibility

• deepagents >= 0.6.0, < 0.7.0
• Python 3.11+