Deep Agents and LangGraph: production deployment overview

Robert Xu from LangChain explains Deep Agents as a harness: the software around a model that manages prompts, tools, memory and execution. He describes its open source components, including filesystem access, planning, subagents and middleware that can intercept model calls.

The talk places Deep Agents above LangChain's tool calling infrastructure and LangGraph's nodes, edges and state. Xu explains how changing an agent's constructor adds the Deep Agents primitives. He also describes using a deep agent inside a deterministic LangGraph workflow, or letting an agent coordinate deterministic flows.

The production discussion focuses on failures during long tasks. Xu says checkpointing lets execution resume near a failed step rather than restart the entire task. He distinguishes session memory from a long-term store for preferences that persist across sessions.

Authentication requires decisions about which tools an AI agent may use and when an action needs additional oversight. The talk also covers interrupts, approvals and progress streaming. Xu closes with a pitch for LangSmith deployments. This is an architecture overview rather than a local installation guide; it does not establish hardware requirements or offline model support.