LangChain and LangGraph

pip install "sovara-langchain[agent]"   # without [agent] if you only use wrap_chat_model

npm install @sovara/langchain @sovara/sdk

Inside a trace, use the middleware for create_agent, or wrap the chat model for anything else (LangGraph nodes, prebuilt agents, structured output). Use one per model. Put the middleware last, so it records the model and prompt that other middleware settled on, and each retry by other middleware as its own step.

import sovara_langchain
from sovara_langchain import SovaraMiddleware, wrap_chat_model

agent = create_agent(model, tools, middleware=[*other_middleware, SovaraMiddleware()])

sovara_langchain.instrument_tools()  # tool calls of agents that don't use the middleware
model = wrap_chat_model(ChatOpenAI(model="gpt-5.4")).bind_tools(tools)

import { sovaraMiddleware } from "@sovara/langchain";
import { wrapChatModel } from "@sovara/langchain/chat-model";
import { instrumentTools } from "@sovara/langchain/tools";

const agent = createAgent({ model, tools, middleware: [...otherMiddleware, sovaraMiddleware()] });
instrumentTools(); // tool calls outside the agent middleware
const wrapped = wrapChatModel(new ChatOpenAI({ model: "gpt-5.4" })).bindTools(tools);

Lessons are one temporary user message at the end of the request; the agent’s messages are never changed. Outside a trace, calls pass through unrecorded, and if Sovara is unavailable they still run. Requires langchain-core 1.x (langchain 1.1+ for the middleware) in Python, or @langchain/core 1.2+ (langchain 1.5+ for the middleware) in TypeScript.

Get started

Agent frameworks

Coding agents

LangChain and LangGraph