Motivation

• Observability — LangSmith
• Rapid Testing (testing prompts, models) — LangGraph Studio
• Evals — LangSmith

Specific Example

• I have a deep research agent in Py using Anthropic native search tool
• Code: https://github.com/langchain-ai/open_deep_research/tree/rlm/expts

Process

• Observability

  • pip install langsmith
  • Add traceable decorator (see ref) to function that you want to trace

  @traceable(run_type="llm",
             name="Overall Research",
             project_name="Anthropic Deep Researcher")
  

• Rapid Testing

  • pip install langgraph
  • Required: Add entrypoint decorator to a function that wraps your agent

  @entrypoint()
  def wrap_agent_with_langgraph(input: dict) -> dict:
      
      # Create and run research agent
      agent = create_research_agent()
  
      result = agent.research(input)
  
      # Return structured result for LangGraph Studio
      return {
          "final_report": result['content'],
          "metadata": result.get('metadata', {}),
          "sources": result.get('sources', [])
      }
  

  • Required: Add LangGraph.json file

  {
    "graphs": {
      "wrap_agent": "./src/open_deep_research/non_langgraph/graph.py:wrap_agent_with_langgraph"
    },
    "python_version": "3.11",
    "env": "./.env",
    "dependencies": [
      "."
    ]
  }
  

  • Then you can run
  • pip install -U "langgraph-cli[inmem]"

  N_JOBS_PER_WORKER=1 langgraph dev
  

  • Optional: Add Configuration
  • Optional: Add typing on input

from open_deep_research.non_langgraph.configuration import Configuration

from typing_extensions import TypedDict

from langchain_core.runnables import RunnableConfig
from langchain_core.messages import convert_to_openai_messages
from langchain_core.messages import AnyMessage

class State(TypedDict):
    messages: List[AnyMessage]

@entrypoint(config_schema=Configuration)
def wrap_agent_with_langgraph(input: State, config: RunnableConfig) -> dict:
    """
    LangGraph entrypoint that wraps the DeepResearchAgent for Studio use.
    
    This function extracts configuration from LangGraph Studio's UI and passes
    it to the research agent, enabling users to customize research parameters
    through the Studio interface.
    
    Args:
        input: Dict containing 'messages' list for the research query
        config: LangGraph RunnableConfig with configurable parameters
        
    Returns:
        Dict with 'final_report' containing the research results
    """
    # Extract LangGraph Studio configuration
    configurable = Configuration.from_runnable_config(config)

    # Convert LangGraph Configuration to ResearchConfig
    research_config = ResearchConfig(
        model=configurable.model,
        writer_model=configurable.writer_model,
        temperature=configurable.temperature,
        max_tokens=configurable.max_tokens,
        max_iterations=configurable.max_iterations,
        max_searches_total=configurable.max_searches_total,
        enable_clarification_qa=configurable.enable_clarification_qa,
        clarification_timeout=configurable.clarification_timeout,
        search_domains=configurable.search_domains,
        search_location=configurable.search_location,
    )

    # Create and run research agent
    agent = create_research_agent(research_config)

    # LangChain Message format to OpenAI
    openai_message_format = convert_to_openai_messages(input['messages'])

    result = agent.research(openai_message_format)

    # Return structured result for LangGraph Studio
    return {
        "final_report": result['content'],
        "metadata": result.get('metadata', {}),
        "sources": result.get('sources', [])
    }

• Evals
  • Run from Studio UI