Agentic design patterns

Reflection

This is the most common way of interacting with an agent. We prompt something, we get a result and we provide feedback, so that the llm can produce better output and so on.

This pattern can also be implemented by creating another agent instructed to provide feedback to the "producing" agent.

• Reflection (Agentic Pattern) can use external feedback from tools to improve output
• Reflection (Agentic Pattern) consistently outperforms direct generation on a variety of tasks
• "LLM as a judge" grading with a rubric gives more consistent results when evaluating Reflection (Agentic pattern)

Tool Use

Agents can be given tools to use to perform different tasks. Tools can be anything from analysis, information gathering, productivity and image processing.

Planning

Agents don't rely on developers to provide the steps to follow, but they come up with their own steps. See Agentic AI can be less or more autonomous. Also Agentic AI works better for well know, linear processes

Multi-agentic workflows

Using a set of agents specialising in a specific thing. They are more difficult to control, because we can't know ahead of time what they'll do.