Deciding Between Skills and Sub-Agents: Key Design Trade-Offs for AI Systems

The question of whether to build a skill or a sub-agent often proves more fundamental than selecting which language model to use. Microsoft Azure lead engineer Kishorekumar Pattabiraman recently published guidance clarifying this architectural choice, arguing that getting the delivery shape right is critical to system success.

The Core Difference#

Skills and sub-agents represent two distinct approaches with incompatible strengths. A skill operates within an ongoing conversation—it can read files, pose clarifying questions, iterate with users, and maintain human oversight throughout execution. A sub-agent, by contrast, accepts a single prompt, executes autonomously to completion, and returns a finished result. Neither design suits the other's use case.

Four Key Decision Dimensions#

Pattabiraman identifies four dimensions that should guide the choice:

• Iteration model: Does the task benefit from back-and-forth refinement, or does it work better as a single autonomous pass?
• Voice fidelity: How important is maintaining a consistent conversational tone throughout the process?
• Human gate placement: Where in the workflow should human decision-makers intervene?
• Task frequency: Is this a one-time custom piece or a repeatable batch operation?

Of these, frequency emerges as the clearest dividing line: one-off, bespoke work tilts toward skills, while repeatable workflows favor sub-agents. The other dimensions require more nuanced trade-offs. Teams must weigh the cost of keeping humans actively engaged in an iterative skill-based loop against the risk of forcing an inherently interactive process into a single autonomous pass that might need correction each time.

Community Perspectives#

Developers across Reddit and Hacker News have highlighted practical distinctions. Some observe that sub-agents start with a clean slate and avoid polluting the shared context window, whereas skills always carry the full conversation history. Others stress that skills shine when reusability across multiple agents or flows matters, while sub-agents make sense when work requires separate context, distinct permissions, or a different knowledge base.