Agentic AI Patterns for Robust AI Applications

Prompt Chaining: LLMs are sequentially linked, where the output of one LLM call feeds the input of the next, useful for tasks like generating structured documents or multi-step data processing.
Routing: An LLM classifies input and directs it to the most suitable specialized task, LLM, or tool, separating concerns and optimizing individual downstream tasks, as seen in customer support systems.
Parallelization: Independent subtasks are run simultaneously, and results are aggregated for speed or enhanced quality, applicable in RAG with query decomposition or analyzing large documents.
Reflection: An agent evaluates its own output against criteria and iteratively refines it based on feedback, useful in code generation or complex problem-solving scenarios.
Tool Use (Function Calling): LLMs interact with the outside world by calling external functions or APIs to fetch data or perform actions, enabling applications like booking appointments or retrieving real-time data.
Planning: A central LLM dynamically breaks down a complex goal into a multi-step plan, delegating execution to worker agents (often using tools), suitable for complex software development or research report generation. According to additional sources, agentic patterns provide structure for designing complex systems, enable modular design, and help manage complexity by offering reusable templates.