Agent Skills for Production LangGraph Agents

Demo from the Medium article: Stop Stuffing Your System Prompt: Build Scalable Agent Skills in LangGraph.

Demonstrates progressive knowledge loading, skill-based domain modularization, and tool-driven skill activation.

Note: This code has evolved beyond the version published with the article. Notable upgrades:

• Introduced a reusable async BaseAgent (graphs/core/) with LLM retry classification (transient vs permanent), tool-call pairing safety, and Langfuse-managed prompts.
• Split skills_agent into typed state + slim nodes, replacing the original monolithic utils/nodes.py.
• The original article described running the agent through Aegra; this repo now runs directly on LangGraph's local dev server for agent testing, with only the graph code and tests kept here.

The Agent Skills concepts in the article still apply; the surrounding implementation has been hardened and simplified.

How it runs

This repo is a LangGraph application. The local code is:

langgraph.json      # LangGraph dev/server configuration
graphs/core/        # reusable BaseAgent
graphs/skills_agent/ # the demo agent + skills
tests/              # unit tests for graphs/

langgraph.json registers skills_agent from ./graphs/skills_agent/agent.py:graph and loads environment variables from ./.env. The graph code remains pure LangGraph: StateGraph, typed state, runtime context, tool nodes, and interrupt/resume behavior live under graphs/.

Quick start

make install
cp .env.example .env  # set OPENAI_API_KEY, or use Ollama defaults
make dev              # runs on http://127.0.0.1:2024

LangGraph Studio is available while the dev server is running:

https://smith.langchain.com/studio/?baseUrl=http://127.0.0.1:2024

Local smoke checks

With make dev running, confirm the graph is registered:

curl -s -X POST http://127.0.0.1:2024/assistants/search \
  -H 'content-type: application/json' \
  -d '{}'

The response should include "graph_id":"skills_agent". For endpoints that accept a graph ID, use skills_agent directly:

curl -s http://127.0.0.1:2024/assistants/skills_agent/graph

Some endpoints require the assistant UUID returned by /assistants/search instead of the graph ID, such as /assistants/{assistant_id}/schemas.

To run a live LLM smoke test through the SDK:

uv run --with langgraph-sdk python - <<'PY'
import asyncio
from langgraph_sdk import get_client


async def main():
    client = get_client(url="http://127.0.0.1:2024")
    thread = await client.threads.create()
    result = await client.runs.wait(
        thread["thread_id"],
        "skills_agent",
        input={
            "messages": [
                {
                    "role": "user",
                    "content": "Say hello and name one skill you can load.",
                }
            ]
        },
    )
    print(result["messages"][-1]["content"])


asyncio.run(main())
PY

Tests

make test

Credits

• LangGraph by LangChain
• n8n Skills Repository by haunchen