Most agent frameworks stop at the ReAct loop: a model, some tools, and a while-loop. That works for a few turns, then the context window fills with tool noise and the agent loses the plot. Deep Agents is a bet that the missing piece isn't a smarter model but better scaffolding around it — the same planning, file system, and delegation tricks that make Claude Code feel general-purpose, extracted and pushed further.
What Sets It Apart
- It is a layer above LangChain's
create_agent, not a from-scratch runtime — so you inherit LangGraph's streaming, persistence, and checkpointing instead of reinventing them. - Context survives long jobs: tool outputs offload to disk, threads summarize, and sub-agents run in isolated context windows so a research detour doesn't pollute the main reasoning.
- The file system is pluggable — local, sandboxed, or remote backends — which means the same agent code can run in a notebook or behind a network sandbox without rewrites.
- Persistent memory with swappable backends gives the agent cross-session recall, turning one-shot scripts into something that accumulates state.
How It Holds Together
The design treats an agent as composable middleware rather than a monolith. Planning, file access, summarization, and human-in-the-loop approval are layers you opt into; custom LangGraph components and MCP servers plug in as sub-agents or tools. A pre-built "Deep Agents Code" variant ships as a worked example of the same primitives aimed at coding.
Who It's For
Great fit if you're building research assistants, coding agents, or any system that needs to grind through multi-step work and remember things between runs, and you already live in the LangChain ecosystem. Look elsewhere if you want a single-turn tool-calling wrapper or a vendor-neutral stack — this is opinionated, LangGraph-shaped, and most valuable exactly when the task is too long for one context window.
