Deep Agents

Deep Agents is LangChain's standalone agent harness library (pip install deepagents) built on LangGraph's durable runtime. Instead of assembling an agent from primitives, you get the whole production stack wired together: an execution environment with tools, a virtual filesystem (in-memory, local disk, LangGraph store, or custom backends) with declarative read/write permission rules, sandboxed shell execution and a QuickJS in-process interpreter, built-in subagents that run in isolated context windows and return a single report, long-term memory via AGENTS.md files, Agent Skills-standard skill loading with progressive disclosure, automatic summarization and context offloading, and human-in-the-loop approval pauses. create_deep_agent(model=..., tools=[...], system_prompt=...) is the entry point; the harness also streams typed events for messages, tools, values, and subagent delegation.

1. Requires Python 3.10+
2. pip install deepagents
3. Set your model provider credentials (any LangChain-supported chat model)
4. Create an agent:

```python

from deepagents import create_deep_agent

agent = create_deep_agent(

model="claude-sonnet-4-5", # any LangChain chat model

tools=[my_tool],

system_prompt="You are a helpful assistant.",

)

result = agent.invoke({"messages": [{"role": "user", "content": "..."}]})

```

1. Optional: pass middleware=[TodoListMiddleware] for task planning, memory=[...] for AGENTS.md files, permissions=[...] to restrict filesystem access