Build a deep research agent by combining various models and search tools
If you want to build an agent that automates deep research on complex topics, this repository is a starting point. It is a framework that lets you connect various LLM providers and search APIs without being tied to a specific cloud or model, allowing you to configure your own research pipeline. You can run it locally, adjust settings as needed, and understand how the agent operates.
Recent updates and performance status
On the 7th of the same month, it was updated to support the GPT-5 model, and GPT-5 results were also reflected in the Deep Research Bench evaluation. On July 30, a blog post was published summarizing the development process from the initial implementation to the current version.
Model and search tool configuration
This tool supports various LLM providers through the init_chat_model() API. You can specify different models for each stage, such as summarization, research, compression, and final report writing. The default is OpenAI's GPT-4.1 series models, but using local models via OpenRouter or Ollama is also possible. The search function uses the Tavily API by default but provides full compatibility with Anthropic and OpenAI's native web search and MCP servers. All settings can be changed directly in the configuration.py file or via the LangGraph Studio UI.
Local execution and deployment methods
After cloning the repository, create a virtual environment and install dependencies using uv. Copy the .env file to set environment variables for models and search tools, then run the local server with the langgraph dev command. The LangGraph Studio UI opens in the browser, where you can enter questions and check the agent's responses. Deployment can be done by hosting via LangGraph Platform or deploying to the Open Agent Platform (OAP) for non-technical users. With OAP, users can directly configure MCP tools and search APIs to customize the agent.
Evaluation benchmarks and costs
Deep Research Bench is a benchmark consisting of 100 PhD-level research tasks created by experts in 22 fields. It uses Gemini, an LLM-as-a-judge method, to compare and evaluate against standard reports written by experts, with leaderboard scores based on RACE scores. Running all 100 examples may incur costs ranging from approximately $20 to $100, depending on the model selection.
Implementation choices and considerations
In addition to the current version, the src/legacy/ folder contains previous implementations based on workflows and multi-agents. The workflow approach emphasizes human-in-the-loop planning stages and sequential processing, focusing on accuracy, while the multi-agent approach uses a structure where multiple researchers work in parallel to increase speed. Although the current implementation is superior in terms of performance, the previous implementations help in understanding the diversity of deep research approaches. Before use, you must ensure that the selected model supports structured output and tool calling, and you must always consider the expected cost and token consumption when running evaluations.
