agent-constructionlisted

# Agent Construction Skill ## When to activate - Designing a multi-agent system with Claude (orchestrator + subagents) - Building a Claude-powered agent that uses tools across multiple turns - Designing memory for a long-running agent (in-context vs. external) - Handling agent errors, retries, and stopping conditions - Implementing agent handoffs between specialized subagents ## When NOT to use - Single-turn Claude API calls — the Claude API skill is sufficient - Simple chatbots without tool use or autonomous decision-making - LangChain/LlamaIndex abstractions — address the abstraction layer directly ## Instructions ### Agent architecture patterns **Single agent with tools** — one Claude instance, multiple tools, loops until task is done: ``` User → Agent → [Tool A] → [Tool B] → Agent → User ``` **Orchestrator + subagents** — one parent spawns specialized children: ``` User → Orchestrator → [ResearchAgent] → [WriterAgent] → Orchestrator → User ``` **Pipeline** — agents hand off results in sequence: ``` User → Agent1(classify) → Agent2(extract) → Agent3(generate) → User ``` Choose the simplest architecture that solves the problem. Single agent with tools handles most cases. ### Tool design ```python # Tool definitions for multi-turn agent loop tools = [ { "name": "search_web", "description": "Search the web for current information. Use when you need facts not in your training data.", "input_schema": { "type": "object",