LangGraph (v2)

AgentV provides native support for LangGraph, allowing you to evaluate state-aware agentic graphs and loops.

1. Setup Your Agent API

Expose your LangGraph agent via a standard HTTP endpoint.

from langgraph.graph import StateGraph
from fastapi import FastAPI

app = FastAPI()

# ... (your graph definition) ...


@app.post("/execute_task")
async def execute(request: dict):
    # Process turn with state
    state = {"input": request["task_description"]}
    result = graph.invoke(state)
    return {"action": "final_answer", "summary": result["output"]}

2. Run Evaluation

Use the langgraph protocol to connect the harness to your agent service.

agentv run \
  --path scenarios/loan_scenario.json \
  --protocol langgraph \
  --agent langgraph://localhost:8000/execute_task \
  --agent-name "LangGraph-Fintech-Orchestrator"

3. Visual DNA Debugging

The langgraph:// protocol is optimized for the Visual Debugger. It captures the internal state transitions between graph nodes and visualizes them as atomic events in the trajectory map.

# Launch the console to see the graph transitions
agentv console