frankxai/claude-skills-library

Overview

This skill teaches how to build production-ready multi-agent systems using OpenAI AgentKit and the Agents SDK, with practical patterns for orchestration, handoffs, routines, and monitoring. It focuses on 2025 best practices: lightweight specialized agents, explicit handoffs, observable workflows, and using the Agents SDK (not legacy Swarm) for production.

How this skill works

The skill inspects agent design, routine definitions, handoff mechanics, and end-to-end orchestration patterns, then maps them to concrete implementation and testing strategies. It covers triage, sequential, and parallel decomposition patterns, guidance on tool assignment, state/context passing, performance optimizations, and production observability. Code snippets and test strategies demonstrate how to implement handoffs, routines, and monitoring using the Agents SDK.

When to use it

• Building customer-facing workflows that need clear routing and specialist handling
• Designing multi-step pipelines where each step requires distinct expertise
• Breaking complex problems into parallel subtasks with a synthesis step
• Migrating legacy educational Swarm setups to production-ready Agents SDK
• Optimizing latency, cost, and observability for agent-driven systems

Best practices

• Keep agents small and single-responsibility with clear system prompts
• Define explicit handoff triggers, context payloads, and return paths
• Give each agent only the tools it needs; avoid full tool sets on every agent
• Design routines with entry/exit conditions, error handling, and success metrics
• Test agents in isolation, then integration and end-to-end scenarios; add observability and tracing early

Example use cases

• Customer service triage routing requests to refund, sales, or technical specialists
• Research pipelines: research_agent → analysis_agent → report_agent sequential flow
• Parallel analysis of large datasets with subtask agents and a synthesis agent
• Refund processing routine with verification, calculation, reversal, and confirmation steps
• Migration project replacing Swarm experiments with Agents SDK-based orchestrations