Azure and Community Guidelines on Choosing Between a Skill or a Sub-Agent

In a recent Azure Architecture blog post, Kishorekumar Pattabiraman, an Azure lead engineer, discusses the factors to consider when deciding between implementing a skill or a sub-agent for building AI systems. Pattabiraman emphasizes the importance of reusability, simplicity, and long-term maintainability in the decision-making process.

The fundamental distinction between a skill and a sub-agent lies in their architecture: a skill operates within an ongoing conversation, while a sub-agent performs a single task independently and provides a final output. Pattabiraman notes that both have their place, depending on the specific task at hand.

To determine the appropriate choice, Pattabiraman highlights four key dimensions to evaluate: iteration model, voice fidelity, human gate placement, and the frequency of task repetition. The frequency of task repetition emerges as the key differentiator, with one-off tasks favoring a skill-based approach and repeatable batch jobs aligning better with a sub-agent.

Other factors, such as the level of human involvement and the complexity of orchestration, require further consideration. Pattabiraman outlines the trade-offs and potential pitfalls associated with each dimension, providing valuable insights for architects and developers in the AI space.