NemoClaw Knowledge Wiki

Langchain DeepResearch

4 min read

Langchain DeepResearch

https://www.youtube.com/watch?v=agGiWUpxkhg This video provides a detailed overview and demonstration of the LangChain Deep Research Agent, an open-source, highly configurable agent built on the LangGraph framework. Here’s a summary of the video’s content: 1. Introduction & Initial Demonstration (0:00-0:56)

  • The speaker introduces the LangChain Deep Research Agent, emphasizing its configurability and open-source nature.
  • Demo: The user asks the agent to plan the “cheapest trip” for three people from New York to Amsterdam and Norway (departing Sept 12, returning Sept 21).
  • The agent immediately asks clarifying questions (e.g., number of travelers, preferred airports, specific Norwegian cities, order of visits).
  • The user provides the additional context (2 roommates, no preferred airports, Amsterdam first, then Oslo, Bergen, Flåm).
  • The agent confirms the details and kicks off the deep research process.

2. Deep Research Agent Architecture (1:07-8:30) The architecture is divided into three main phases: Scope, Research, and Write.

  • Phase 1: Scope (2:09-4:07) User Clarification: The agent (using structured output from an LLM) can ask follow-up questions to gather necessary context from the user, especially when the initial prompt is vague. This ensures the research is well-targeted. Brief Generation: After clarifying, the entire chat history (user requests and agent clarifications) is compressed into a single, cohesive “research brief” using an LLM. This helps the agent focus and iterate against a clear target, preventing context window bloat. LangSmith Trace Example (3:02): The video demonstrates the clarify_with_user step checking if clarification is needed and the write_research_brief step condensing the conversation into a research_brief string.

  • Phase 2: Research (4:07-8:30) Research Supervisor: This component takes the research brief and spawns “research sub-agents” to research specific sub-topics in-depth and in parallel. This is a context engineering technique to break down complex research into manageable, isolated tasks. Research Sub-agents: Each sub-agent focuses on a single sub-topic within its own separate context window. They utilize a tool-calling loop, iteratively calling various tools (e.g., tavily_search, Search_SearchGoogle, Search_SearchHotels, Search_OneWayFlights, Search_RoundtripFlights, and custom MCP tools) to gather information. Findings Compression: Before returning to the supervisor, each sub-agent performs an additional compression step to synthesize its raw tool outputs and findings into a clean, comprehensive mini-report. This is crucial because raw tool outputs can be messy and contain irrelevant information, potentially overloading the supervisor’s context window. The compression step emphasizes preserving all relevant information verbatim rather than summarizing. Supervisor Decision: The supervisor continuously receives compressed findings from the sub-agents, analyzes them, and decides whether to spawn more sub-agents for further research or conclude the research if sufficient information has been gathered. This allows for maximal flexibility and dynamic research depth, mimicking human research patterns. LangSmith Trace Example (5:10): The trace shows the supervisor kicking off three parallel research topics (flights, intra-European transport, accommodation). It then dives into one sub-agent’s tool-calling loop, showing calls to various search tools, and finally the ResearchComplete tool and the compress_research step before returning findings.

  • Phase 3: Write (8:30-9:30) One-Shot Report Generation: Once the supervisor is satisfied, all condensed research findings from the various sub-agents are presented to a “writer” LLM in a single request. This “one-shot” approach for final report generation was found to produce more cohesive and less disjointed reports compared to parallel report writing. LangSmith Trace Example (8:58): The trace for the final report shows the entire research brief and the compressed findings from all sub-agents being fed into a single LLM call to generate the final report.

3. Running Locally & Open Agent Platform (9:30-12:46)

  • Local Setup: The agent is open source and built on LangGraph. The video shows how to clone the GitHub repository (open_deep_research), create a virtual environment, install dependencies, set API keys in a .env file (OpenAI, Tavily, Google, LangSmith tracing), and run LangGraph Studio locally (uv run langgraph dev).
  • LangGraph Studio UI: The Studio UI provides a visual representation of the agent’s graph, allowing users to interact with it, test different configurations (e.g., allow clarification, max concurrent research units, search API type, different LLM models for summarization, research, and final report), and observe traces.
  • Open Agent Platform: For users who don’t want to run the code locally, the agent is also available on the Open Agent Platform, providing an easy-to-use UI for configuration and interaction.
  • Final Report Display (11:24): The video concludes by showing the fully generated trip planning report in the Open Agent Platform UI, featuring: An overview of the trip. Cheapest flights (outbound, intra-European, return) with detected itineraries, prices, and direct booking links. Intra-European transport options (trains, buses, flights) between cities. Cheapest accommodation options for each city with direct booking links. A recommended sample itinerary. Key tips for lowering costs further. A comprehensive list of official operator booking links (sources).
  • The speaker encourages viewers to explore the GitHub repository and an accompanying blog post for more details on the design decisions.

Graph View