Agentic workflows use LLMs as autonomous agents to achieve a goal. The LLM is invoked iteratively to initiate and manage processes. For example, it can autonomously handle tasks such as scheduling meetings, processing customer queries, or even conducting research by interacting with APIs and databases. The model uses contextual understanding to navigate these tasks, making decisions based on the information it processes in real-time. Any automated workflow that invokes an AI agent is considered “agentic”, even if part or most of the workflow is executed as traditional software. This is because special considerations must be made to accommodate the unique requirements of AI agents, no matter how much of the workflow they automate. The key characteristics of an agentic workflow include:
  • Autonomy: The LLM operates independently for extended periods, adapting to dynamic environments and making real-time adjustments based on the evolving context of the task.
  • Contextual understanding: The model maintains an understanding and memory of the ongoing context and uses this information to guide its actions, ensuring coherent and consistent responses.
  • Decision-making: The LLM makes decisions based on the information it processes, selecting appropriate strategies and adapting to challenges to achieve its goals.
  • Interaction with external systems: The model can interact with APIs, databases, and other tools to gather information, perform computations, or execute actions, extending its capabilities beyond its inherent knowledge and skills.
Rather than single-shot prompt engineering, agentic workflows can be enhanced through the application of flow engineering techniques, which involve designing and optimizing the workflow itself to guide the agent’s decision-making process and improve the quality of its outputs. This seeks to maintain a balance of autonomy and structure in the agent’s operations.