// the find
langchain-ai/deepagents
The batteries-included agent harness.
Deep Agents is an opinionated agent harness layered on top of LangGraph that bundles filesystem access, sub-agents, context management, and loadable skills so you don't have to wire them yourself. It sits between LangChain's minimal create_agent and raw LangGraph — the FAQ actually explains this clearly, which is rare for a framework at this abstraction level. Aimed at developers who want a working long-horizon agent in an afternoon without writing their own planning loop.
The layered-stack explanation in the FAQ is genuinely useful — it tells you when to drop down to LangGraph instead of pretending Deep Agents is always the right answer. Context management (thread summarization, offloading tool outputs to disk) tackles a real problem that most agent harnesses ignore until your context window explodes mid-task. The Skills abstraction is pleasantly simple: SKILL.md files the agent reads on demand, no custom DSL, no runtime magic. The security model is honest — 'trust the LLM, enforce at the sandbox level' — rather than hand-waving about model self-policing.
You're two abstraction layers deep (your code → Deep Agents → LangGraph → LangChain), and LangGraph has a history of API churn; a LangGraph breaking change lands on you whether you use it directly or not. The SKILL.md approach has no schema or validation — if skills are ever user-supplied or pulled from untrusted sources, that's a prompt injection surface with no guardrails in sight. There's no story for deterministic or auditable behavior: if your production workflow needs predictable step counts or bounded retries, the harness defaults are fully LLM-driven and reining that in means dropping back to raw LangGraph anyway. The repo also ships 'Deep Agents Code', a terminal coding agent that competes with Claude Code and Cursor — it's a different product target muddying what this repo actually is.