WorkflowArena benchmarks runtime workflow orchestration by LLM agents
Original title: WorkflowArena: Benchmarking Executable Workflow Orchestration in LLM-Based Multi-Agent Systems
On a hard subset, Claude Opus 5.5-authored workflows substantially improved weaker executors with the executor held fixed.
WorkflowArena evaluates whether language-model agents can decide how to delegate and author executable workflows at run time. It contains 145 long-horizon tasks across eight professional domains, each with a workspace, an interdependent deliverable, and a hidden verifier. Tests of nine models found no configuration best in every domain. Claude models in Claude Code used workflows less often but with larger worker groups, while GPT models in Codex delegated more often in smaller units.
- The benchmark has 145 tasks spanning eight professional domains.
- Delegation is recorded so quality and work organization can be scored together.
- Nine models were tested; none led in every domain.
- Claude in Claude Code used larger worker groups and fewer workflows at high effort.
- GPT models in Codex delegated more frequently in smaller units.
- Products
- WorkflowArenaClaude CodeCodex
- AI models
- ClaudeGPTClaude Opus 5.5
From the original article
1 min read at arxiv.orgLarge language model agents are increasingly deployed as multi-agent systems rather than as a standalone model operating within a single context. Recent agent harnesses allow the model to instantiate worker agents and to author the executable workflow that coordinates them, shifting the design of system structure from developers to the model at run time.
By Zheng Jiang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.