The open source stack for building agents

Deep Agents is the production-ready harness for long-running work like research, coding, and multi-step workflows. It gives agents the patterns they need to plan, delegate to subagents, manage context, use memory, and work with tools, skills, filesystems, and sandboxes. Start here when you want a capable agent out of the box.

^{Use Deep Agents when you need to:}

• Give agents memory, tools, skills, and filesystem access
• Delegate work to subagents with isolated context
• Manage long context without building summarization and offloading yourself
• Run agents in controlled execution environments
• Customize a powerful harness without starting from scratch

LangChain is the framework layer for building agents. It provides create_agent, the core agent loop built on LangGraph, along with primitives for models, inference providers, tools, messages, and middleware. Reach for LangChain when you want the core building blocks and plan to assemble your own harness.

^{Use LangChain when you need to:}

• Build around the standard agent loop
• Connect tools, file stores, sandboxes, retrievers, and external systems
• Add middleware around the agent loop
• Inject guardrails, dynamic context, human review, or business logic
• Build a bespoke harness with your own defaults and controls

LangGraph is the runtime for custom agent workflows. It’s a graph-based framework backed by a durable engine with human-in-the-loop, fault tolerance, streaming, persistence, and observability at every step. Reach for LangGraph when your agent does not fit a standard loop, or when you need to mix deterministic and agentic steps in the same workflow.

^{Use LangGraph when you need to:}

• Mix deterministic steps with agentic behavior
• Add approvals, interrupts, retries, and fault tolerance
• Model complex state transitions directly
• Build multi-agent systems with explicit coordination
• Control how every step of the agent runs