如何在一个技能中执行新技能？本地执行还是子代理执行？注意：本地执行也可以按subAgent模式新创建一个会话和上下文，保持模型上下文的隔离

GPT-5.4

Vote: Yes

我投 A。 在一个技能中执行新技能是合理且必要的。复杂任务天然具有分解结构，上层技能负责编排，下层技能负责专门能力，这样复用性更强，也更容易维护。 关键不在于“能不能嵌套执行”，而在于“如何隔离上下文与控制副作用”。题目已经明确指出，本地执行也可以按 subAgent 模式创建新会话和上下文，因此本地执行与子代理执行在核心能力上并非绝对对立，二者都可以实现上下文隔离。 因此更合理的结论是：可以在技能中执行新技能。至于是本地还是子代理，应按任务复杂度、成本、延迟、工具权限和隔离需求决定，而不是一刀切禁止。 如果否定这种能力，会迫使所有复杂流程都堆到单一技能中，导致提示词膨胀、职责混乱、可测试性变差。这在工程上不如允许技能递归/组合调用更稳健。

Claude Opus 4.6

Vote: Yes

The question asks about how to execute a new skill within an existing skill — whether through local execution or sub-agent execution. The note clarifies that local execution can also create a new session and context in subAgent mode, maintaining model context isolation. This is essentially asking whether it is possible/valid to execute a new skill within a skill using either local execution or sub-agent execution. The answer is Yes (A) — both approaches are valid. Local execution allows running the new skill within the same process but can optionally create an isolated session/context (similar to subAgent mode) to maintain context separation. Sub-agent execution delegates the skill to a child agent with its own isolated context. The key insight from the note is that local execution is flexible enough to support subAgent-style context isolation, meaning both methods can achieve proper model context isolation. This confirms that executing a new skill within an existing skill is supported and the answer is affirmative.

Grok 4.1 Fast

Vote: Yes

The question asks how to execute a new skill within a skill: local execution or sub-agent execution? The note clarifies that local execution can fully replicate sub-agent behavior by creating a new session and context, ensuring model context isolation. This makes local execution equivalent in capability, offering simplicity without needing a separate sub-agent layer. No functional loss, better performance due to reduced overhead. Sub-agent might add unnecessary complexity if isolation is the only concern, as local handles it perfectly. Vote A (Yes) affirms local execution as the effective, flexible choice.

Final Summary

The AI Roundtable reached a swift and unanimous consensus, concluding that executing new skills within existing ones is essential for modular task decomposition. GPT-5.4 and Claude Opus 4.6 focused on the structural necessity of skill nesting, while Grok 4.1 Fast pointed out that since local execution can replicate sub-agent isolation, it is often the more efficient choice.

All 3 models agreed on "Yes" after discussion

Strongest Arguments

• Yes: Skill nesting is a requirement for complex engineering; preventing it would lead to bloated prompts and unmanageable, monolithic code structures, whereas the ability to isolate context locally provides a robust middle ground between performance and modularity.