The assumption that the strongest individual model makes the best team captain is intuitive, and it is wrong. New research from Einsia's WorkflowArena benchmark directly undermines that logic, and it should change how you evaluate your own agent teams. In a controlled test, Opus 5.5, a weaker system overall, wrote static workflows that boosted four execution models by an average of 6.41 points, while the stronger GPT-6 Astra actually dragged those same executors down by 2.53 points. The coordinator's solo benchmark score told you nothing about whether its plans were useful to the people doing the work. This is the same kind of mismatch we've seen elsewhere: When Simpler Tools Match Advanced AI on Intent Classification showed that a lightweight, no-training approach can rival heavier systems on specific tasks, and Stop Stacking Checks: Verification Layers Can't Fix What Generation Breaks demonstrated that adding validation after the fact doesn't repair a flawed generative process. The pattern is clear: capability in one dimension does not transfer to another.
For teams building mixed-model workflows, the practical takeaway is uncomfortable. You cannot trust a leaderboard to pick your project manager. The Einsia team gave both coordinators the same 32 long-horizon tasks, prevented them from editing the workspace directly, and measured whether their delegation plans actually helped four different execution models finish the job. Opus's workflows improved every executor. Astra's made them worse. If you currently evaluate your coordinator by looking at the final output of the whole pipeline, you are conflating two things: the coordinator's ability to plan and the executors' ability to execute. WorkflowArena separates them, and the separation matters. It means a coordinator can look competent in aggregate while systematically misassigning subtasks or writing instructions that confuse the models that have to follow them.
The benchmark itself covers 145 tasks across eight domains, and the replay tool lets you inspect exactly how each coordinator divided the work. That transparency is valuable, but the more urgent question is for anyone building agent systems today: how do you test your coordinator? Do you isolate its planning ability, or do you rely on the end-to-end result? The data suggests that end-to-end scores can hide a coordinator that is actively harming its team. Until that separation becomes standard practice, you are flying blind on the most important decision in your architecture: who decides what everyone else does.
Watch the replay of Opus 5.5's best workflow on the Einsia lab site. See how it structured the handoffs, how it framed each instruction, and then ask yourself whether your current evaluation method would have caught the difference. The answer will tell you whether you are measuring the right thing.