launchdarkly-ai-langchain-agents

LangChain handler for launchdarkly-ai-server using LangGraph's StateGraph (langgraph). Delegates the full agentic loop to the LangGraph ReAct agent. Works with any BaseChatModel — defaults to ChatOpenAI.

provides_for: ['*', 'agent'] — matches any flag variation where meta.mode is "agent" and no more-specific handler is registered. LangChain is a framework adapter, not a provider: it routes through langchain-anthropic, langchain-openai, and others at runtime based on config.provider.name. Use '*' so that flags configured with provider.name = "Anthropic" or "OpenAI" are automatically handled without requiring a separate native handler.

Installation

pip install launchdarkly-ai-server launchdarkly-ai-langchain-agents

The default model is ChatOpenAI, so set OPENAI_API_KEY unless you pass a custom BaseChatModel.

Usage

With the default model (ChatOpenAI)

import asyncio
from launchdarkly_ai_server import config, shutdown
from launchdarkly_ai_langchain_agents import create_langchain_agents_handler

async def main():
    result = await config(
        key="my-ai-config-flag",
        handler=create_langchain_agents_handler(),
        tool_handlers={"search": lambda q: "..."},
    ).invoke(
        "Research and summarize feature flagging best practices",
        {"kind": "user", "key": "user-123"},
    )

    print(result.response)
    await shutdown()

asyncio.run(main())

Tool handlers may be sync or async. Sync handlers run on the event-loop thread, so blocking I/O stalls the agent. Keep graph __handoff_* handlers synchronous: routing records the selected edge on the call itself, and moving them onto asyncio.to_thread would break that.

With a custom BaseChatModel

from langchain_anthropic import ChatAnthropic
from launchdarkly_ai_langchain_agents import create_langchain_agents_handler

handler = create_langchain_agents_handler(ChatAnthropic(model="claude-opus-4-5"))

A constructed instance cannot see flag parameters. Pass a function instead if the model should be built after evaluation:

handler = create_langchain_agents_handler(
    lambda config: ChatAnthropic(
        **{
            **(config["model"].get("parameters") or {}),
            "model": config["model"]["name"],
        }
    )
)

Convenience wrapper

import asyncio
from launchdarkly_ai_langchain_agents import langchain_agents

async def main():
    user_input = "Research feature flagging best practices"
    result = await langchain_agents(
        user_input,
        {"kind": "user", "key": "user-123"},
        {"key": "my-ai-config-flag"},
        variables={"user_input": user_input},
    )
    print(result.response)

asyncio.run(main())

Agent graphs — langchain_graph()

Runs a LaunchDarkly agent graph with the LangChain agent handler pre-bound. Equivalent to calling the base graph() with handlers=[create_langchain_agents_handler()]. See the core client docs for the full graph() API.

import asyncio
from launchdarkly_ai_langchain_agents import langchain_graph

async def main():
    result = await langchain_graph("support-graph").invoke(
        "I was double charged",
        {"kind": "user", "key": "user-123"},
    )
    print(result["response"])

asyncio.run(main())

Native graph adapter — to_lang_graph()

Converts a resolve_graph() result into a framework-native LangGraph StateGraph. Pre-order traversal (root → leaves) builds a compiled StateGraph. Single-child edges become direct edges after a tool loop; multi-child edges use Command-returning handoff tools (bound with parallel_tool_calls=False) so the model picks exactly one target.

import asyncio
from launchdarkly_ai_server import resolve_graph
from launchdarkly_ai_langchain_agents import to_lang_graph

async def main():
    ctx = {"kind": "user", "key": "user-123"}
    result = await to_lang_graph(
        resolve_graph("support-graph", context=ctx),
        {
            "tool_handlers": registry.tools,
            "context": ctx,
            # optional: supply your own model per node
            "model_factory": lambda node: ChatOpenAI(model=node["config"]["model"]["name"]),
        },
    ).invoke("I was double charged")
    print(result["response"])

asyncio.run(main())

How It Works

• Uses the system prompt and conversation history defined in your LaunchDarkly flag config.
• Template placeholders ({{variable}}) in the prompt are substituted using variables before the call.
• The LangGraph ReAct agent manages the full reasoning and tool-call loop autonomously — reasoning through steps, calling tools, and deciding when to stop.
• Emits an OTel span and LaunchDarkly telemetry for every call.

Choosing Between langchain-agents and langchain-messages

Environment Variables