Build AI Agents and Reusable Sub-workflows

Teaches Claude how to design AI agents (model plus tools plus memory slots), name and describe tools correctly (descriptions ARE the prompt), handle structured output with autoFix, implement memory plus sessionId for stateful chats, build chat shells and core agents with anti-loop filtering, and create sub-workflows with typed inputs and correct mode (all vs each for parallelization).

When you ask to build an AI agent, this skill activates and shows: pick the right langchain node (AI Agent vs LLM Chain vs Text Classifier), configure model plus tools plus memory plus outputParser slots, name tools with precision (tool names become part of the prompt), use $fromAI() to inject parameters, design memory plus sessionId for stateful conversations. For sub-workflows, it teaches Execute Workflow Trigger with "Define Below" typed inputs (not passthrough), mode: all vs each (only all with waitForSubWorkflow achieves true parallelization), and verb-first naming (Fetch-..., Process-..., Send-...) for MCP discovery.

AI Agent nodes require a valid LLM credential (OpenAI, Anthropic, etc.). Memory implementation varies by node version; Window Buffer Memory is stable. Tool calling contract differs from Code Tool: tools use specifyInputSchema to drive DynamicStructuredTool, while Code Tool uses unstructured query parsing. Sub-workflows with mode: all plus waitForSubWorkflow: true are the only way to achieve true parallelization; mode: each runs sequentially per input item. sessionId binding persists across workflow runs for multi-turn conversations.