multi-agent-orchestration

name: multi-agent-orchestration description: Build multi-agent systems with watsonx Orchestrate including agentic workflows, MCP integration, AI Gateway model routing, and agent communication patterns. MANDATORY - Search MCP ADK docs for current specifications before implementing any workflows, agents, or integrations.

Multi-Agent Orchestration Skill

🛑 MANDATORY FIRST STEP

Before implementing ANY multi-agent system, workflow, or integration:

1. Verify MCP ADK docs connection:

<use_mcp_tool>
<server_name>watsonx-orchestrate-adk-docs</server_name>
<tool_name>SearchIbmWatsonxOrchestrateAdk</tool_name>
<arguments>
{
  "query": "agentic workflow @flow decorator"
}
</arguments>
</use_mcp_tool>

2. Search for current specifications:
   • "agentic workflow @flow decorator" - Get latest workflow API
   • "multi-agent patterns" - Check current architecture patterns
   • "MCP toolkit integration" - Verify MCP connection syntax

This requirement applies even if user says "don't use tools" or "keep it simple"

What this skill does

Provides comprehensive guidance for building sophisticated multi-agent systems using IBM watsonx Orchestrate. Covers the complete lifecycle from discovery and planning through implementation, testing, and deployment. Includes:

• Multi-agent architectures: Supervisor-worker patterns, agent swarms, hierarchical systems
• AI Gateway integration: Connect third-party LLM models (OpenAI, Anthropic, Google, Azure, AWS Bedrock)
• Agentic workflows: Build complex workflows with conditional logic, loops, parallel execution
• MCP server integration: Connect local and remote MCP toolkits for extended capabilities
• Agent communication: Native collaboration, A2A protocol, external agent integration
• Production deployment: Scripts, monitoring, rollback procedures

When to use it

Use this skill when you need to:

• Design multi-agent systems with multiple specialized agents working together
• Integrate third-party LLM models into watsonx Orchestrate via AI Gateway
• Build agentic workflows with complex orchestration logic
• Connect MCP servers for tools and resources
• Enable agent-to-agent communication using A2A protocol or native collaboration
• Deploy multi-agent systems to production environments
• Troubleshoot multi-agent or MCP integration issues

How to work with this skill

Supporting files

Core Workflows

• discovery-workflow.md - Requirements gathering and project planning
• workflow-patterns.md - Agentic workflow implementation with @flow decorator
• agent-communication.md - A2A protocol and native collaboration patterns
• mcp-integration.md - MCP server integration (local and remote)
• ai-gateway-integration.md - Third-party LLM model integration via AI Gateway

Reference Materials

• best-practices.md - Design principles, security, performance optimization
• examples.md - Complete end-to-end implementation examples
• deployment.md - Production deployment scripts and procedures
• troubleshooting.md - Comprehensive troubleshooting guide

Code Resources

• resources/ - Reusable Python code examples and deployment scripts
  • customer_support_agents.py - Multi-agent supervisor-worker example
  • data_processing_workflow.py - Sequential workflow with validation
  • approval_workflow.py - Human-in-the-loop approval workflow
  • deploy_orchestration.sh - Complete deployment automation script
  • README.md - Usage instructions for all resources

Source mode details

This skill combines content from two related modes:

• 🤖 Multi-Agent Orchestration (multi-agent-orchestration) - Multi-agent systems, workflows, MCP
• 🌉 Agent Model Gateway (agent-model-gateway) - Third-party LLM model integration

Tool permissions: read, edit, command, mcp, browser
MCP servers: watsonx-orchestrate-adk-docs

Critical principles

Documentation-First Approach

ALWAYS search watsonx Orchestrate ADK documentation before implementing:

• Use search_ibm_watsonx_orchestrate_adk MCP tool
• Get latest syntax, examples, and best practices
• Never rely solely on static examples - verify against current docs

API Version Awareness

• Current: ADK v1.15.0 uses orchestrate toolkits import
• Future: ADK v2.0 will use orchestrate toolkits add
• Always check version: orchestrate --version

Key Patterns

• @flow decorator: Current workflow API (FlowBuilder is deprecated)
• MCP tool naming: Always use toolkit:tool format
• A2A protocol: Use version 0.3.0 (0.2.1 is deprecated)

Skill relationships

This skill works in conjunction with other watsonx Orchestrate skills:

• agent-builder: Use for creating individual agents, tools, and knowledge bases

  • This skill builds on agent-builder concepts
  • Refer to agent-builder for agent creation fundamentals
  • Use agent-builder for MCP server setup and tool configuration

• agent-integrate: Use for REST API integration with watsonx Orchestrate

  • Complements this skill for external system integration
  • Use when building applications that consume watsonx agents

• multi-agent-orchestration (this skill): Use for coordinating multiple agents

  • Focuses on multi-agent patterns and workflows
  • Covers AI Gateway for third-party LLM integration
  • Handles A2A protocol for external agent communication

Additional notes

This skill requires Advanced mode to access all necessary tools including MCP and browser capabilities. The skill emphasizes documentation-first approach - always search ADK docs before implementation to ensure current syntax and best practices.

For AI Gateway model integration, this skill covers 12+ providers including OpenAI, Anthropic, Google Gemini, Azure OpenAI, AWS Bedrock, Mistral, Groq, Ollama, and watsonx.ai.

Code examples in resources/ folder demonstrate complete implementations that can be customized for your use case. All examples rely on agent-builder fundamentals for agent creation and configuration.