Comparison of Deep Agents, LangChain, and LangGraph

Deep Agents, LangChain, and LangGraph are open-source agent stacks designed to allow developers to directly own models, context, and execution methods. The three layers can be combined with each other; as you move from runtime to harness, the level of abstraction increases and the scope of control decreases.

• LangGraph: A graph-based runtime for constructing custom agent workflows. It is built on a durable execution engine and provides features such as human-in-the-loop.
• LangChain: An agent framework and abstraction/integration layer. It provides a minimal agent structure where an LLM loops and calls tools, and middleware can be used to add deterministic steps to the execution loop.
• Deep Agents: An off-the-shelf agent harness that provides context engineering by default. It embeds various best practices for delivering appropriate context to the model at the right time.

Deep Agents includes a filesystem for reading and writing information outside the LLM context window, subagents for isolating specialized tasks, skills that provide instructions and scripts loaded on demand, and memory to support learning across executions. Using create_deep_agent, you can create an agent by specifying the model, tools, system prompt, and skill paths.

LangChain's default agent operates around a simple LLM-tool calling loop. When behavior needs to be extended, such as summarizing before the context fills up or running a validator at the end, you can utilize middleware hooks. Deep Agents is described as a form of this LangChain agent combined with multiple middlewares.