LangGraph Adapter

DriftGuard ships a first-class LangGraph adapter that creates a review node you can drop directly into any StateGraph.

Install dependencies

pip install "driftguard-ai[demo]"

Basic usage

from driftguard import DriftGuard
from driftguard import make_langgraph_review_node

guard = DriftGuard()
review_node = make_langgraph_review_node(guard)

Drop into a StateGraph

from langgraph.graph import StateGraph, END
from typing import TypedDict

class AgentState(TypedDict):
    candidate_action: str
    guard_warnings_count: int
    guard_confidence: float
    guard_top_warning: dict | None

graph = StateGraph(AgentState)

graph.add_node("plan", plan_node)
graph.add_node("review", review_node)
graph.add_node("execute", execute_node)

graph.set_entry_point("plan")
graph.add_edge("plan", "review")
graph.add_conditional_edges(
    "review",
    lambda state: "execute" if state["guard_warnings_count"] == 0 else "revise",
    {"execute": "execute", "revise": "revise"},
)
graph.add_edge("execute", END)

app = graph.compile()

State keys written by the review node

Key	Type	Description
guard_review	RetrievalResponse	Full review object
guard_warnings_count	int	Number of warnings found
guard_confidence	float	Highest warning confidence
guard_top_warning	dict or None	trigger, risk, confidence

Custom key names

review_node = make_langgraph_review_node(
    guard,
    action_key="next_action",         # read from this state key
    review_key="dg_review",           # write full review here
    warnings_key="dg_warnings",
    confidence_key="dg_confidence",
    top_warning_key="dg_top_warning",
)

Policy inside the node

review_node = make_langgraph_review_node(
    guard,
    policy="block",
    min_confidence=0.8,
)

Demo agent

A full LangGraph demo with an LLM planner is included:

export OPENAI_API_KEY=your-key
python demo/langgraph/demo_agent.py --duration-seconds 120 --step-delay 4 --reset-graph

This runs a real planner → DriftGuard review → revision → execution loop with a live OpenAI-compatible model.