Open Brain
arXivPaperNeeds Review

WorkflowArena: Benchmarking Executable Workflow Orchestration in LLM-Based Multi-Agent Systems

WorkflowArena evaluates how agents organize delegated work as well as the final deliverable. Its results suggest orchestration quality can affect task performance independently of the executing model.

arXiv · Nov 9, 2026
Open Source Open MarkdownOpen JSON
Source Summary

WorkflowArena includes **145 long-horizon tasks** across **8 professional domains**. Agents choose whether and how to delegate; hidden verifiers assess integrated deliverables, while the benchmark records delegation. The study tests **nine models**.

Practical Implication

Evaluate your agent’s work breakdown and delegation alongside task completion. In a challenging subset, workflows authored by Claude Opus 5.5 improved weaker executors when the executing model was held fixed.

Agent-Ready Context
WorkflowArena includes **145 long-horizon tasks** across **8 professional domains**. Agents choose whether and how to delegate; hidden verifiers assess integrated deliverables, while the benchmark records delegation. The study tests **nine models**.

Evaluate your agent’s work breakdown and delegation alongside task completion. In a challenging subset, workflows authored by Claude Opus 5.5 improved weaker executors when the executing model was held fixed.

No single configuration led across all domains. The reported delegation patterns differ between Claude Code and Codex at high effort, but the paper does not establish that either strategy is generally better.
Connected Context · Feed7 Judgment

This benchmark makes delegation decisions and integrated deliverables part of agent evaluation across long, multi-domain tasks. Its findings favor evaluating workflow quality alongside completion, but the lack of a configuration that leads across domains limits any general ranking of delegation strategies.

Context Map
benchmarkcodingresearch#multi-agent#agent-evals
Uncertainty
No single configuration led across all domains. The reported delegation patterns differ between Claude Code and Codex at high effort, but the paper does not establish that either strategy is generally better.