Multi-Agent Orchestration Workflow
Pattern Selection
| Pattern | When to Use | Complexity | RAG Query |
|---|---|---|---|
| Delegation | Route to specialists by task type | Low | "agent delegation routing" |
| Sequential | Pipeline with handoffs | Low | "sequential agent pipeline" |
| Parallel | Independent concurrent tasks | Medium | "parallel agent execution" |
| Hierarchical | Manager coordinates workers | Medium | "hierarchical agent manager" |
| Mesh | Peer-to-peer collaboration | High | "agent mesh communication" |
Framework-Specific Patterns
ADK Multi-Agent
Uses sub_agents for delegation.
RAG Query: mcp__agentic-rag__query_sdk("sub_agents delegation", sdk="adk", mode="build")
Key: Agent description drives routing decisions.
LangGraph Multi-Agent
Uses graph nodes for each agent, conditional edges for routing.
RAG Query: mcp__agentic-rag__query_sdk("multi-agent graph", sdk="langgraph", mode="build")
Key: State must flow between agent nodes.
CrewAI Multi-Agent
Uses Crew with agents and tasks.
RAG Query: mcp__agentic-rag__query_sdk("crew agents tasks", sdk="crewai", mode="build")
Key: Process type (sequential vs hierarchical) changes everything.
OpenAI Multi-Agent
Uses handoffs between agents.
RAG Query: mcp__agentic-rag__query_sdk("agent handoff", sdk="openai", mode="build")
Key: Handoff descriptions determine routing.
Routing Strategy Decision
| Strategy | Best For | Tradeoff | RAG Query |
|---|---|---|---|
| Keyword | Simple, predictable tasks | Brittle | "keyword routing" |
| LLM-based | Complex, nuanced routing | Slower, costlier | "llm routing classification" |
| Semantic | Similar intent detection | Needs embeddings | "semantic routing embedding" |
| Rule-based | Deterministic, auditable | Manual maintenance | "rule based routing" |
Critical Gotchas
These cause multi-agent failures:
- •Vague descriptions - Agents won't be routed correctly if descriptions are generic
- •Missing state passing - Context lost between agents
- •Circular delegation - Agent A → B → A infinite loop
- •No termination condition - Loops forever
- •Sequential when could parallel - Unnecessary latency
- •No error handling - One agent failure crashes system
State Sharing Patterns
| Pattern | When to Use | RAG Query |
|---|---|---|
| Shared dict/TypedDict | Simple, synchronous | "shared state dict" |
| Message history | Conversational | "message passing agents" |
| External store | Persistence needed | "agent state persistence" |
| Event bus | Decoupled agents | "event driven agents" |
Workflow: Building Multi-Agent System
Step 1: Define Agent Responsibilities
Each agent needs:
- •Clear, specific role
- •Well-defined capabilities
- •Distinct from other agents
Step 2: Choose Orchestration Pattern
Use table above to select pattern based on your workflow.
Step 3: Design Routing Logic
RAG Query: mcp__agentic-rag__search("[routing strategy] [framework]", mode="build")
Step 4: Implement State Sharing
RAG Query: mcp__agentic-rag__search("state sharing [framework]", mode="build")
Step 5: Add Termination Conditions
- •Max iterations
- •Goal achievement check
- •Timeout
Step 6: Error Handling
- •Fallback agents
- •Retry logic
- •Graceful degradation
Common Error Patterns
| Symptom | Likely Cause | Fix |
|---|---|---|
| Wrong agent called | Vague description | Make descriptions specific |
| Infinite loop | No termination | Add max iterations, goal check |
| Context lost | No state passing | Implement state sharing |
| Slow response | Sequential unnecessary | Use parallel where possible |
| One failure crashes all | No error handling | Add try/catch, fallbacks |
Performance Optimization
| Issue | Solution | RAG Query |
|---|---|---|
| Latency | Parallel where possible | "parallel agent optimization" |
| Token usage | Shorter agent contexts | "agent context optimization" |
| Routing errors | Better descriptions | "agent description best practices" |
| State bloat | Prune old messages | "state management cleanup" |