LangChain Open Deep Research Internals: A step-by-step guide (www.bolshchikov.com)

LangChain published a deep, step-by-step walkthrough of Open Deep Research that peels back the architecture and runtime state changes of its multi-agent research pipeline. Rather than a high-level overview, the guide traces execution from initial user scoping through brief generation, a supervisor that reflects and spawns research sub-agents, and a Reporter that composes final outputs. It’s significant because it documents design patterns and state management techniques you don’t usually see—helping practitioners build robust, scalable research agents that avoid context bloat and enable parallel, dynamic reasoning. The post highlights two core patterns: a reflection pattern (LLMs critique and iterate on their own outputs) and a manual tool-orchestration pattern where “tools” are merely schemas bound to the LLM and executed manually by the framework. This lets the supervisor spawn entire subgraphs (research sub-agents) in parallel, track per-subgraph state (messages, topics, tool call IDs), cap iterations to avoid loops, and compress/summarize large results before returning lightweight confirmations to the main state. Key nodes and tools include think_tool, conduct_research, research_complete, compress_research, and web_search (Tavily). The trade-off is losing automatic tool execution in exchange for fine-grained control over routing, memory, and orchestration—crucial for complex, recursive multi-agent workflows and producing large artifacts without blowing up the LLM context.