Deep Agents

Deep Agents is an open source agent harness from LangChain for teams that want a ready-to-run agent loop instead of wiring every part by hand. The project describes itself as a batteries-included harness: it ships with planning defaults, sub-agents, context handling, filesystem access, shell access, persistent memory, human approval flows, skills, and a tool interface that can also call MCP servers. It is built on LangGraph, so builders get checkpointing, streaming, state, and the same runtime path used by larger LangChain agent deployments. The fastest path is the Python package. The README shows `uv add deepagents`, then `create_deep_agent(model=..., tools=..., system_prompt=...)`. From there, the agent can plan a task, read and write files, call custom tools, delegate parts of the job to sub-agents with their own context windows, and keep long tool outputs out of the chat thread when they should live on disk. The key design choice is that Deep Agents is opinionated but not closed. You can override or replace pieces without forking the project. Deep Agents is useful for developers building research agents, coding agents, internal operations agents, and long-running assistants that need more than a single prompt and a few function calls. It fits teams that already use LangChain or LangGraph, but it is also approachable for builders who want sensible defaults before they learn the lower-level graph primitives. Model choice is flexible. The README says it works with tool-calling models from frontier APIs, open-weight providers, and local stacks through LangChain chat model integrations. The feature set is strongest when the agent needs durable context. Sub-agents help split tasks without stuffing every detail into one prompt. The filesystem layer lets the agent place notes, drafts, and tool outputs where future steps can reuse them. Human-in-the-loop controls let teams review sensitive tool calls before execution. Persistent memory adds cross-session recall when the application needs it. For production work, the project points users to LangSmith for tracing, evaluation, monitoring, and deployment support. Pricing is straightforward because Deep Agents is open source under the MIT license. There is no license fee for the harness itself. Users still pay for whatever model provider, runtime, tracing product, storage, or deployment platform they connect to it. That makes the main buying decision technical: if you want a light single-agent wrapper, plain LangChain may be enough. If you want a custom state machine, use LangGraph directly. If you want the full agent harness with planning, delegation, context handling, file tools, and production hooks already assembled, Deep Agents is the practical starting point. One small but important advantage is that Deep Agents keeps the path from prototype to production fairly direct. A team can begin with the default harness, add a few internal tools, inspect behavior through traces, and then move lower in the LangGraph stack only when the workflow needs a custom control path. That makes the project useful both as a real library and as a readable pattern for designing durable agent systems.