MS‑Eval measures when main-agent → sub-agent delegation actually helps (and when it just costs money)

Teams building agent stacks must decide whether a central “main” agent should sequentially manage tools or delegate sub-tasks to specialized sub-agents. The paper MS‑Eval (arXiv) formalizes that decision problem: Main-Agent Delegation and Sub-Agent Execution (MAD‑SAE) can reduce context bloat and enable parallel work, but delegation quality and orchestration overhead are rarely measured systematically. What MS‑Eval introduces and reports MS‑Eval is a unified evaluation kit that combines an MCP‑intensive benchmark (multi‑sub‑task, tool‑heavy scenarios), a controlled execution harness, tailored delegation/orchestration metrics, and a composite MSQ index for aggregate assessment. The authors evaluate four MAD‑SAE variants across eight contemporary models. Key findings reported by the paper: Sub-agent specialization (Spec) is empirically the best practical strategy among the tested delegation variants. Open-source models are largely competitive with closed models on these tasks, with some exceptions. MAD‑SAE generally raises operational cost; its payoff is contingent — it improves performance and efficiency for many open-source models but often increases cost and degrades performance for some closed-source models. The authors’ case study attributes gains to three mechanisms: stronger single‑agent solutions (better per-subtask accuracy), effective context offloading (reduced prompt/context size at the main agent), and true parallelism across sub-agents. Conversely, unnecessary or redundant delegation converts extra cost into