WorkflowArena: best model orchestrates multi-agent workflows to only 36.95% full solves

机器之心 · wechat · 2026-11-10

Navers Lab (Einsia AI) released WorkflowArena, a benchmark for multi-agent orchestration ability: 145 long-horizon tasks across 8 domains (software engineering, business ops, science, math, finance, law, content design, 3D), evaluating 25 reasoning configurations of 9 frontier models. Models get a blank orchestration console and are scored only on final deliverables meeting all acceptance criteria.

Key findings

Top system Claude Opus 5.5 (max) fully solves just 36.95%; formal math is near-saturated (95.83%), but software engineering tops out at 29.41%, finance 18.75%, law 9.09%; no system wins everywhere
Two orchestration styles: Claude models run few but large workflows (0.8 calls/task, 5-6 sub-agents each); GPT models run frequent small ones (2 calls/task, <2 sub-agents)

Orchestration is a separate skill: swapping only the orchestrator, Opus-written workflows lift Kimi K3.1 Pro from 35.75 to 46.61; yet the stronger GPT-6 Astra as orchestrator hurts executors — being good at tasks ≠ being good at orchestrating.

Economics: per-task costs differ by tens of times; spending more tokens doesn't buy accuracy — orchestration quality sets the ceiling.

Original post →

WorkflowArena: best model orchestrates multi-agent workflows to only 36.95% full solves

More from coding & agent