Deep Agents: LangChain architecture and deployment overview

Sydney Runkle and Jake Broekhuizen explain Deep Agents, LangChain's open source harness for complex, long-running work. They discuss research and coding tasks that need more support than a basic model-and-tool loop, including planning, subagent delegation and context summarization.

The file system discussion focuses on keeping information accessible without putting everything in the prompt. Agents can load relevant files when needed and offload large tool results. Runkle describes local files as useful for development, then explains custom backends and a composite backend that can combine data sources. She cautions that summarization loses information, which is one reason to reduce context size before compacting it.

For production, the speakers cover human approval, durable execution and streaming through LangGraph. They describe LangSmith tracing and an assistant for questions about individual traces and failures across runs. Their starting advice is to write specific prompts, clarify tool descriptions and set up evaluations early.

Sandbox placement brings a tradeoff: running the whole AI agent inside it puts API keys there too, while isolating tool execution adds network latency. The discussion also leaves room for deterministic workflows, with custom middleware for checks and Deep Agents as one step in a larger workflow.