LLM application architecture should match problem complexity. Pattern 1: Direct Prompt-Response. When to use: well-defined task, single-turn, reliable output format. Implementation: input validation -> prompt template -> LLM call -> output parsing -> response. Examples: document summarization, sentiment classification, text translation. Pattern 2: RAG (Retrieval-Augmented Generation). When to use: factual questions requiring current or specialized knowledge, source attribution needed. Implementation: embed query -> vector search -> rerank -> augmented prompt -> LLM -> parsed response with citations. Examples: enterprise Q&A, customer support with knowledge base. Pattern 3: Tool-Using Agent. When to use: task requires external data retrieval or actions, multi-step reasoning needed. Implementation: agent loop (reason -> select tool -> execute -> observe -> repeat until done). Examples: research assistant, code debugging, data analysis. Pattern 4: Multi-Agent. When to use: tasks benefit from parallel processing, specialized expertise, verification. Implementation: orchestrator assigns tasks to specialized agents, collects results, synthesizes. Examples: complex research reports, software development pipeline. Decision framework: Start with simplest architecture, add complexity only when simpler fails. Agents add 3-10x latency and cost - justify with quality improvement. Multi-agent adds coordination overhead - usually only for tasks >30 minutes of work.