LangSmith/Deep Agents

• Deep Agents

  • langchain-ai/deepagents — 29,853 GitHub/Stars as of 2026-09-29 Tue
  • "The batteries-included agent harness." Built on langgraph.
  • Written in Python; a TypeScript version lives at langchain-ai/deepagentsjs.
  • Built-in capabilities: a virtual filesystem for context management, AI/Agent/Subagent spawning, long-term memory in AGENTS.md files, AI/Agent/Skills, task planning, and AI/Workflow/Human in the Loop approval.

  • Virtual filesystem

    • Every deep agent gets file tools: ls, read_file, write_file, edit_file, delete, glob and grep. With a sandbox backend it also gets execute for shell commands.
    • The files live wherever the configured backend puts them:
      • StateBackend — in the agent's langgraph state, scoped to one thread; the default
      • FilesystemBackend — real files under a root directory on local disk
      • LocalShellBackend — local disk plus shell execution
      • StoreBackend — the LangGraph store, persisting across threads, with storage scoped per user, per assistant or per thread
      • CompositeBackend — routes path prefixes to different backends, such as /memories/ to the store and everything else to state
      • a custom backend that implements ls, read, write, edit, glob and grep
    • Permission rules declare which paths the agent can read or write. They don't apply to sandbox backends.
    • The filesystem does the context offloading:
      • a tool result over 20,000 tokens is saved to a file and replaced by its path and first 10 lines; the agent re-reads or greps it when needed
      • once context passes 85% of the model's window, older large file writes and edits in history are replaced by a pointer to the file
    • Source: Deep Agents overview, Backends, Context engineering