菜单
Use when starting work on any LangChain / LangGraph / DeepAgents project. Entry point for the develop -> middleware -> evaluate -> deploy lifecycle, mapping each step to the right official tool.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when editing a DeepAgents project — adding tools, sub-agents, modifying the system prompt, choosing a filesystem backend, or composing extra middlewares (retries, fallbacks, HITL) on top.
Use when creating a new LangChain / LangGraph / DeepAgents project from scratch. Picks the right scaffolder for graphs vs. DeepAgents vs. LCEL chains.
Use when productionising or deploying a LangChain / LangGraph / DeepAgents agent. Covers durable execution (checkpointers, thread_id), the production middleware stack, three deploy targets (LangSmith Cloud, Cloud Run, Docker), secrets, scaling, and post-deploy verification.
Use when editing a non-agentic LCEL pipeline — composing Runnables, retrievers, embeddings, chat models, parsers, or building a RAG chain. For agents (LLM + tools loop), use the middleware skill instead.
Use when building a custom-graph LangGraph agent — when `create_agent(...)` + middleware isn't enough and you need explicit StateGraph control (multi-graph workflows, custom routing, non-standard state schemas, parallel branches). For the common case, use `create_agent` first (see middleware skill).
Use when authoring eval datasets, writing evaluators, running evals against a LangChain / LangGraph / DeepAgents project, comparing eval results between runs, or writing unit/integration tests for an agent.
| name | langchain-agents-workflow |
| description | Use when starting work on any LangChain / LangGraph / DeepAgents project. Entry point for the develop -> middleware -> evaluate -> deploy lifecycle, mapping each step to the right official tool. |
This skill is the entry point — read it first, then load the more specific skill for the step you're on.
langchain >= 1.2 (v1 series — middleware-first create_agent)langgraph >= 1.1, langgraph-cli >= 0.4deepagents >= 0.5.3 (async sub-agents, structured sub-agent responses, filesystem permissions; model=None deprecated)langsmith >= 0.7 (pytest plugin + new evaluator API)langchain-anthropic >= 1.4, langchain-openai >= 1.0If a project pins anything below these floors, suggest the bump before writing code — the API shapes in this bundle assume the v1+ surface.
This skill bundle pairs with the mcpdoc MCP server. The two have distinct roles; use both.
mcpdoc (MCP server) | This skill bundle | |
|---|---|---|
| Purpose | Live API reference | Opinionated playbook |
| Content | Whatever's on docs.langchain.com right now | How to think about LangChain projects |
| When to use | "What's the signature of SummarizationMiddleware?" / "What kwargs does create_agent take?" / "What import path for X?" | "What's the production middleware stack?" / "How do I wire Cloud Run + Secret Manager + Postgres checkpointer together?" / "Which mistakes does an agent typically make here?" |
| How to use | Call fetch_docs(url) or list_doc_sources() | Load skills based on description triggers |
| Drift risk | Zero — always live | Owner updates as ecosystem evolves; some rot tolerated |
Rule of thumb: when you need an exact API detail, fetch from mcpdoc. When you need to make a design decision, load a skill. When in doubt, do both.
| Goal | Skill |
|---|---|
| Start a new agent project | langchain-agents-scaffold |
| Build a modern agent (most cases) | langchain-agents-middleware ← first; uses create_agent(...) with middleware |
| Add nodes/edges/tools to a custom LangGraph | langchain-agents-langgraph-code |
| Customize a DeepAgent | langchain-agents-deepagents-code |
| Build a non-agentic LCEL pipeline (chains, RAG) | langchain-agents-langchain-code |
| Write or run evals; unit/integration test agents | langchain-agents-langsmith-evals |
| Deploy + productionise | langchain-agents-deploy |
| Debug / read traces / OTEL | langchain-agents-observability |
Three layers in the modern stack:
create_agent(model, tools, middleware=...) — the v1 default for building agents. Middleware is how you add retries, fallbacks, summarization, HITL, PII handling, call limits. Read the middleware skill for the production stack.StateGraph) — drop down when create_agent isn't enough (multi-graph workflows, custom state, parallel branches). Read langgraph-code.create_deep_agent(...) is create_agent(...) pre-loaded with FilesystemMiddleware + SubAgentMiddleware + TodoListMiddleware. Read deepagents-code.For non-agentic flows (RAG, classification), use plain LCEL chains — middleware does not apply to chains.
| Stage | Command(s) |
|---|---|
| Scaffold a LangGraph project | langgraph new my-agent --template react-agent |
| Scaffold a DeepAgent / chain | No scaffolder — write a small agent.py (see scaffold skill) |
| Install deps | pip install -e . or uv sync |
| Iterate on a graph | langgraph dev |
| Run an agent ad hoc | python -c "from agent.agent import agent; print(agent.invoke({'messages': [...] }))" |
| Run evals | python evals/run.py |
| Unit-test agents (no API calls) | pytest with LLMToolEmulator middleware |
| Deploy to LangSmith Cloud | langgraph build -t my-agent && langgraph deploy |
| Deploy to Cloud Run | gcloud run deploy my-agent --source . |
| Deploy as a Docker image | docker build && docker run --env-file .env |
mcpdoc, don't guess. If mcpdoc isn't configured, ask the user to set it up (see this repo's README) before you write LangChain code.pip install — pip show langchain langgraph deepagents langsmith..env contents — refer to keys by name only.ls, tree -L 2) before assuming layout.LANGSMITH_API_KEY — for tracing and evals.LANGSMITH_TRACING=true — enables trace capture.LANGSMITH_PROJECT — trace bucket name.OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.If any are missing when needed, fail fast with a clear message that names the missing variable.