Agentic Workflow Patterns

Agentic workflows are specific patterns of behaviours for agents to complete their tasks. The diagrams below, taken from Weaviate’s What Are Agentic Workflows? Patterns, Use Cases, Examples, and More blog post show three agentic workflow design patterns which are Planning, Tool Use, and Reflection. Each pattern can be applied to agentic solutions of the same categories to maximise the effectiveness of the agents.

Planning Pattern

The Planning Pattern empowers AI agents to autonomously break down complex tasks into simpler subtasks, enhancing reasoning and accuracy by reducing LLM cognitive load. This is particularly useful for ambiguous tasks requiring adaptive problem-solving, such as debugging. While effective for complex reasoning, planning introduces unpredictability and should be reserved for tasks demanding significant problem-solving and multi-step reasoning.

Tool Use Pattern

As previously mentioned in the section Why Use Agents?, LLMs are limited by their reliance on static training data, leading to potential inaccuracies. Retrieval Augmented Generation (RAG) addresses this by incorporating external, real-time data for improved context. The Tool Use Pattern surpasses simple RAG by enabling dynamic interaction with the real world, allowing agents to utilise external resources like APIs, web browsers, and code interpreters for specific tasks, thus enhancing their capabilities and enabling them to perform actions beyond simple data retrieval.

Reflection Pattern

Reflection is a self-feedback mechanism that significantly enhances agentic workflows by enabling agents to iteratively evaluate and refine their outputs. The Reflection Pattern is particularly effective for tasks requiring multiple attempts, such as code generation, where agents can use execution errors to improve their output. By incorporating self-critique into their memory, agents achieve continuous improvement, adapt to user preferences, and enhance problem-solving efficiency without direct human intervention.

TL;DR

By following best practices in the form of agentic workflow patterns, we can optimise our agentic solutions.