LangChain's FilesystemBackend Lets Deep Agents Read and Write Real Disk Files

LangChain's Deep Agents support a FilesystemBackend that allows agents to read and write actual files on a user's machine, unlike StateBackend which stores files temporarily within a LangGraph conversation thread. Files created by the agent — such as Markdown or Python files — persist on disk even after the script finishes running. A key distinction is that thread_id does not isolate files when using local disk storage, meaning multiple threads can access the same real folder. Developers can use CompositeBackend to keep internal agent artifacts separate from actual project files. Because the agent can modify real operating-system files under a configured directory, the backend must be used with care.