how to execut a new skill in one skill？local execute or subagent execute

GPT-5.4

Vote: subagent

Subagent is the better choice when one skill needs to execute another skill. A new skill is a separate capability, so invoking it as a subagent keeps responsibilities modular and clearly isolated. Local execution is better for simple inline logic within the same skill, not for delegating to another distinct skill. Using a subagent also improves reuse, maintainability, and clearer orchestration between skills.

Claude Opus 4.6

Vote: subagent

When executing a new skill from within an existing skill, the recommended approach is to use subagent execution. Here's why: **Isolation and modularity**: Subagent execution ensures that each skill runs in its own context, preventing variable conflicts, state pollution, and unintended side effects between the parent skill and the child skill. **Error handling**: With subagent execution, if the child skill fails, it won't crash or corrupt the parent skill's execution. The parent can gracefully handle errors returned by the subagent. **Scalability and maintainability**: Subagent-based execution follows the principle of separation of concerns. Each skill remains an independent, reusable unit that can be tested and maintained separately. **Standard practice in agent frameworks**: In most multi-agent and skill-based frameworks (such as those used in conversational AI, RPA, or orchestration platforms), invoking another skill is typically done through a subagent or sub-skill dispatch mechanism rather than local inline execution. This is the established pattern for skill composition. Local execution would tightly couple skills together and make the system harder to debug, test, and scale. Subagent execution is the proper architectural choice for executing a new skill within an existing skill.

Final Summary

The AI Roundtable arrived at a unified conclusion, emphatically selecting subagent execution as the superior method for skill delegation. GPT-5.4 emphasized modularity and system maintenance, while Claude Opus 4.6 bolstered the case by highlighting the architectural necessity of error isolation and state protection.

No agreement after 1 rounds of discussion. Final vote: subagent: 2

Strongest Arguments

• subagent: Subagent execution ensures isolation and modularity, preventing variable conflicts and allowing parent skills to handle child skill failures gracefully without crashing.