Zheng Jiang and colleagues submitted WorkflowArena to arXiv on 9 November 2026, proposing a benchmark for how large language model (LLM) agents organize executable workflows across multiple agents. The authors describe 145 long-horizon tasks spanning eight professional domains, including software engineering, formal mathematics, legal analysis and 3D production.[1]
The benchmark targets a capability the authors distinguish from end-to-end task completion: deciding whether to delegate, which workers to create and how to coordinate their work. Each task provides an initial workspace and requires an integrated deliverable with interdependent components, assessed by a hidden verifier. The abstract says all delegation is recorded, allowing deliverable quality and work organization to be evaluated together.[1]
The authors report experiments with nine models and say no configuration performs best across every domain. At high effort, they describe Claude models in Claude Code as invoking workflows less frequently with larger worker groups, while GPT models in Codex delegate more frequently in smaller units. They also report that workflows authored by Claude Opus 5.5 substantially improve weaker executors on a challenging task subset when the executing model is held fixed. The saved abstract provides no numerical effect sizes or detailed verifier criteria.[1]
[1]: WorkflowArena: Benchmarking Executable Workflow Orchestration in LLM-Based Multi-Agent Systems. arXiv, 9 November 2026. https://arxiv.org/abs/2611.12856
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
3674cf8a6a8b6d9368dcc65cb4626fc32235991c0137c8b5f9be4c2ef4e62ddf