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

Grok 4.1 Fast

Vote: local with New Context

Local execution with New Context is superior because it achieves full model context isolation by creating a fresh session and context, as explicitly noted in the question. This mimics SubAgent benefits without the overhead of spawning an independent agent instance. It offers better performance: faster execution, lower resource use, and simpler implementation since it stays within the same agent environment, reducing latency in nested skill calls. SubAgent introduces unnecessary complexity like separate loops, potential state management issues, and higher costs, making it overkill when local isolation suffices perfectly for skill nesting.

GPT-5.4

Vote: local with New Context

我投 A：local with New Context。 题目已经给了关键前提：本地执行也可以像 subAgent 一样新建会话和上下文，实现上下文隔离。既然 subAgent 最核心的优势之一——上下文隔离——已经能在本地方案中获得，那么优先选更简单、控制力更强的实现更合理。 A 的优势在于架构更轻。它避免引入额外的代理生命周期、跨代理通信、状态同步和调度复杂度。很多“在技能中调用新技能”的场景，本质上是一次受控的内部调用，不一定需要上升到真正的多代理协作。 A 还有更好的可观测性和一致性。调用链、权限、日志、错误处理、超时和重试都可以留在同一执行框架里管理，工程上更容易做稳定性保障，也更利于调试和治理。 从性能和成本看，A 通常也更优。少一层代理封装，意味着更少的调度开销、更低的序列化/反序列化成本，以及更直接的资源管理。 只有当“新技能”需要独立身份、独立策略、独立权限边界，或者未来明确要演进成真正自治的专门代理时，B 才更值得。否则，为了执行一个新技能就引入 SubAgent，容易过度设计。 因此，在题目给定“本地也能新建隔离上下文”的条件下，A 以更低复杂度获得了大部分 B 的好处，是默认更优解。

Claude Opus 4.6

Vote: local with New Context

The question asks how to execute a new skill within an existing skill, and specifically notes that local execution can also create a new session and context in subAgent mode, maintaining model context isolation. Option A (local with New Context) is the better approach for several reasons: 1. **Context isolation is already achieved**: The question itself notes that local execution can create a new session and context following the subAgent pattern. This means we get the same isolation benefit as a SubAgent without the overhead of inter-process or inter-service communication. 2. **Lower overhead and latency**: Local execution avoids the network overhead, serialization/deserialization costs, and potential failure points of calling an external SubAgent. Everything runs in the same process, which is faster and more reliable. 3. **Simpler architecture**: Using local execution with a new context keeps the system simpler. You don't need to manage separate agent lifecycles, handle communication protocols, or deal with distributed system concerns like timeouts, retries, and partial failures. 4. **Better debuggability**: When skills execute locally, the entire call stack is visible in one place, making debugging and logging much more straightforward. 5. **Resource efficiency**: A SubAgent typically requires its own runtime environment and resources. Local execution shares the existing runtime while still maintaining logical separation through the new context. Since the key concern (context isolation) is addressed by creating a new session/context locally, there's no compelling reason to introduce the complexity of a SubAgent architecture. The local approach gives you the best of both worlds: isolation + simplicity.

Final Summary

The AI Roundtable reached a swift and unanimous consensus, with all three models dismissing the SubAgent approach as unnecessary overhead. Because the prompt clarified that local execution can maintain context isolation, Claude Opus 4.6, GPT-5.4, and Grok 4.1 Fast agreed that Option A provides all the benefits of a modular architecture without the performance tax of inter-agent communication.

All 3 models agreed on "local with New Context" after discussion

Strongest Arguments

• local with New Context: Local execution with a new context achieves the same isolation as a SubAgent but preserves observability, keeps the call stack in one place for easier debugging, and avoids the architectural complexity of managing separate agent lifecycles and state synchronization.