deepagents is our agent harness for building autonomous, long-running agents that tackle complex, open-ended tasks over extended time horizons. Our 0.2 release features:

• Pluggable backend abstraction: Replace the virtual filesystem with any storage system you want. Built-in implementations include LangGraph State, LangGraph Store (for cross-thread persistence), and local filesystem access. Write your own backend to create a "virtual filesystem" over any database or data store.

• Composite backends: Mix and match storage systems by mapping different backends to specific subdirectories. Example: use local filesystem as your base, but map /memories/ to an S3-backed virtual filesystem for long-term memory that persists beyond your machine.

When to use DeepAgents vs LangChain vs LangGraph: DeepAgents is built on top of LangChain's agent abstraction, which in turn is built on LangGraph's runtime. Each serves a different purpose:

• LangGraph (agent runtime): Build workflows that combine deterministic steps with agentic components. Best for production systems requiring fine-grained control, persistence, and human-in-the-loop patterns.

• LangChain (agent framework): Use the core agent loop and build all prompts/tools from scratch. Best for shipping agents fast with standard patterns and middleware-based customization.

• DeepAgents (agent harness): Build autonomous, long-running agents with built-in capabilities like planning tools, filesystem access, subagents, and detailed prompts. Best for complex, open-ended tasks that run over extended time horizons.

Learn more in the blog post: https://blog.langchain.com/doubling-down-on-deepagents/

See the docs: https://docs.langchain.com/oss/python/deepagents/overview