When does an agent actually need another agent?

First, the intuition.

Delegation is useful when a task can be split into genuine independent work. But every handoff introduces extra messages, duplicated reading, coordination, and potential loss of context. More agents may mean more activity rather than more progress.

What the researchers actually did

MS-Eval studies 166 multi-tool tasks spanning eight real-world scenarios. It separates tasks that could benefit from delegation from those that do not need it, then compares a solo baseline with general-purpose, specialized, and plan-first delegation strategies under a shared harness.

What the results suggest

• Specialized delegation is the strongest practical approach among the tested strategies, but its advantage depends on model and task.
• Delegated runs generally cost more. Some open-weight model setups gained performance, while tested closed-model configurations often spent more without gaining success.
• Over-delegation and redundant subagents are recurring sources of avoidable work.

How convincing is the evidence?

A direct solo-agent baseline uses the same underlying harness, task sets, and tool environment. That improves comparison quality compared with cross-paper anecdotes. The evaluation also reports completion and resource metrics.

What remains uncertain?

Realized tokens, aggregate parallel computation, and tool activity are not fully equalized. The benchmark includes general MCP-intensive tasks rather than only repository-level software engineering. Results should not be generalized to every model or deployment.

Connections worth exploring

• Delegation resembles parallel algorithm design: coordination cost can erase a theoretical speedup.
• A useful routing policy should predict task decomposition value before spawning agents.
• Compare success per dollar, not just total solved tasks.