| name | langgraph-multi-agent-supervisor |
| description | Build LangGraph multi-agent systems with intelligent supervisor orchestration of specialized worker agents (Genie, RAG, MCP, LLM, custom). Use when creating multi-agent systems, implementing supervisor patterns, routing queries to specialized agents, or building hierarchical/sequential/parallel agent workflows. |
LangGraph Multi-Agent Supervisor Skill
Build LangGraph multi-agent systems with a supervisor that orchestrates specialized worker agents. The supervisor intelligently routes user queries to the appropriate agent and manages the overall conversation flow.
Core Architecture
- Supervisor node: LLM-powered routing that selects the best worker agent for each query
- Worker agents: Specialized agents (Genie, RAG, MCP, custom) that handle domain-specific tasks
- State management: Shared state across agents with result aggregation
- Routing strategies: LLM-based (semantic), rule-based (keywords), sequential, or parallel
Supervisor Patterns
Choose the right pattern based on your use case:
| Pattern | Use Case | Pros | Cons |
|---|
| Simple | Single agent selection per query | Easy to implement | No multi-step workflows |
| Hierarchical | Complex domains with sub-domains | Scalable organization | More complex setup |
| Sequential | Multi-step workflows | Results build on each other | Slower execution |
| Parallel | Independent queries needing multiple agents | Faster execution | More complex aggregation |
See references/patterns.md for complete implementation code for all 4 patterns.
Worker Agent Types
| Type | Description | Key Config |
|---|
| Genie | Natural language queries against Databricks Genie spaces | space_id |
| RAG | Document retrieval via Vector Search | index_name, endpoint_name |
| MCP | External tool execution via MCP protocol | tools list |
| LLM | General-purpose LLM responses | model endpoint |
| Custom | Any custom function | User-defined |
See references/worker-agents.md for implementation code and configuration-driven agent creation.
Quick Start
1. Define State and Workers
from typing import TypedDict, Annotated, Sequence
import operator
from langchain_core.messages import BaseMessage, AIMessage
from langgraph.graph import StateGraph, END
from databricks_langchain import ChatDatabricks
class SupervisorState(TypedDict):
messages: Annotated[Sequence[BaseMessage], operator.add]
next_agent: str
agent_results: dict
final_response: str
WORKER_AGENTS = {
"sales_agent": {"description": "Sales data queries", "type": "genie", "space_id": "..."},
"docs_agent": {"description": "Documentation search", "type": "rag", "index_name": "..."},
}
2. Build Graph
graph = StateGraph(SupervisorState)
graph.add_node("supervisor", supervisor_node)
for name, config in WORKER_AGENTS.items():
graph.add_node(name, create_worker_node(name, config))
graph.add_edge(name, END)
graph.set_entry_point("supervisor")
graph.add_conditional_edges("supervisor", route_to_agent, routes)
agent = graph.compile()
3. Enable MLflow Tracking
import mlflow
mlflow.langchain.autolog()
supervisor = create_supervisor_agent()
Best Practices
- Agent design: Keep agents focused and specialized with clear descriptions for routing
- Routing: Use semantic routing (LLM) for flexibility, add rule-based for known patterns
- State: Keep shared state minimal, use agent-specific state for details
- Performance: Cache results, use parallel execution when possible, implement timeouts
- Observability: Enable MLflow tracing, log routing decisions, track execution time
References
