LLM agents are increasingly embedded in real-world workflows, yet they must cope with diverse domain‑specific tools and complex multi‑sub‑task problems. Sequential tool invocation leads to inefficiency and severe context bloat. The Main‑Agent Delegation and Sub‑Agent Execution (MAD‑SAE) paradigm mitigates these issues by assigning sub‑tasks to specialized sub‑agents that run in isolated contexts in parallel. However, the quality of delegation and orchestration has not been quantitatively measured. To fill this gap we introduce MS‑Eval, a unified evaluation kit that includes an MCP‑intensive benchmark, a controlled execution harness, dedicated delegation and orchestration metrics, and the comprehensive MSQ index. Experiments on four MAD‑SAE variants across eight frontier models yield three main findings: (1) sub‑agent specialization (Spec) is the most practical practice; (2) open‑source models are largely competitive with only a few exceptions; (3) MAD‑SAE generally raises cost but delivers both performance and efficiency gains for most open‑source models, while closed‑source models experience higher cost and degraded performance. A case study traces these gains to better single‑agent solutions, effective context offloading, and genuine sub‑task parallelism, whereas unnecessary or redundant delegation turns the same cost into pure overhead.
Code repository: https://github.com/LINs-lab/MS-Eval
Review