Deep Agents

Deep Agents is an opinionated, batteries-included agent harness built on LangGraph that ships with sub-agent spawning, a pluggable filesystem, shell access, context/memory management, and human-in-the-loop approval out of the box. It is designed for long-horizon, multi-step agent work where a minimal agent loop isn’t enough.

Built on top of LangChain’s create_agent, it is model-agnostic (works with any tool-calling LLM, frontier or open-weight) and production-ready, with first-class tracing, evaluation, and deployment support via LangSmith, while every default piece (filesystem backend, sub-agent behavior, memory store) can be overridden or replaced without forking.

What You Get

• create_deep_agent() factory that assembles a full agent with sensible production defaults
• Sub-agent delegation with isolated context windows for parallel or specialized tasks
• A pluggable filesystem (local, sandboxed, or remote backends) the agent can read/write/search
• Middleware for context summarization, tool-output offloading, and long-thread management
• Human-in-the-loop hooks to approve, edit, or reject tool calls before execution

Common Use Cases

• Building long-horizon research or coding agents that need to plan across many steps
• Adding sub-agent delegation to isolate context for specialized sub-tasks
• Giving an agent safe, auditable shell and filesystem access via pluggable backends
• Deploying production agents that need tracing, evaluation, and observability via LangSmith
• Prototyping agent behavior with reusable ‘skills’ the agent can load on demand

Under The Hood

Architecture - Deep Agents is a monorepo (libs/deepagents, libs/code, libs/cli, libs/acp, libs/evals) where the core deepagents package wraps LangChain’s create_agent with a stack of middleware/ modules (subagents, filesystem, memory, summarization, permissions, skills) and pluggable backends/ (local shell, LangGraph store, LangSmith, filesystem) selected at agent construction time via create_deep_agent(). Tech Stack - Built on langchain/langchain-core and LangGraph’s graph runtime for streaming/persistence/checkpointing, with optional extras for AWS (langchain-aws), sandboxed JS execution (langchain-quickjs), and video (av/pillow); a sibling deepagents.js package provides the same harness for JavaScript/TypeScript. Code Quality - Roughly 78 test files across the deepagents package exercise middleware behavior, backends, and agent construction, and the repo uses release-please for automated versioned releases across its multiple libs, indicating a structured release and test discipline for a fast-moving project. API Design - create_deep_agent(model=..., tools=[...], system_prompt=...) gets a fully working agent running in a few lines, and every subsystem (filesystem, sub-agents, memory) is overridable independently, though the breadth of middleware and backend options means going beyond defaults requires reading the architecture docs to know which piece to swap.