WorkflowArena

WorkflowArena is a benchmark announced on 10 November 2026 by Einsia AI's Navers Lab, with Tsinghua, for whether LLM-based multi-agent systems can divide work, coordinate workers, and revise the workflow that ties them together. The official account @EinsiaAI called it a benchmark for autonomous multi-agent workflow orchestration on real, long-horizon work. The same day's lab research listing and project page at lab.einsia.ai/workflowarena are the public entry points; the paper is arXiv:2611.12856 (submitted 9 November 2026), and the code is at Einsia/WorkflowArena.

According to the paper abstract, existing benchmarks either score end-to-end task completion or evaluate coordination under structures specified in advance. WorkflowArena treats workflow orchestration as its own capability. It has 145 long-horizon tasks across 8 professional domains and 17 task forms, including software engineering, formal mathematics, legal analysis, and 3D production. Each task starts from an initial workspace and asks for an integrated deliverable made of interdependent parts, scored by a hidden, task-specific verifier that includes integration checks — correct pieces that are not reconciled need not pass.

An agent may finish the work alone or use a Workflow tool to launch up to 16 concurrent workers. Workers are fresh instances of the same model–harness–effort configuration, share the workspace, and every delegation is recorded so the main agent can inspect results and revise the workflow. A four-hour wall-clock budget is shared by the main agent and the workers, so organizing the work consumes time that could have been spent doing it. The paper reports experiments on 25 model–effort configurations from nine models: Claude Opus 5.5 in Claude Code at max effort recorded the highest Overall Resolved, 36.95%, and no configuration led every domain. In a transfer test that held the executor fixed, static workflows authored by Opus scored 45.04 versus 29.69 for ones authored by GPT-6 Astra when the executor was Kimi K3.1 Pro. Those figures come from the paper; we have not independently reproduced them.

Highlights:

• Official project page: https://lab.einsia.ai/workflowarena/ ;
• Announcement post (2026-11-10T14:30:01Z): https://x.com/EinsiaAI/status/2120162063905681632 ;
• Paper (arXiv:2611.12856): https://www.alphaxiv.org/abs/2611.12856 ;
• Code: https://github.com/Einsia/WorkflowArena ;
• This is the name of an evaluation benchmark, not the earlier Quant-Arena benchmark from the same lab (quant research).

Why it matters

Agent harnesses already let a model spawn workers at run time and write the executable workflow that coordinates them, so the quality of that organization is a distinct capability. Most public benchmarks still ask whether the task was finished, or they score coordination under a structure the benchmark author fixed in advance. WorkflowArena asks whether the model itself decides how to split the work, what context each worker gets, and how results are checked and combined, and it uses hidden verifiers on the integrated deliverable.

It is a different eval from Quant-Arena, which the same lab published earlier for quantitative-research strategies. The 36.95% headline, the transfer-test scores, and the finding that no configuration led every domain are the authors' reported results. Overall Resolved is the equal-weight mean of the eight domain resolved rates, with a separate continuous score for partial progress, so those numbers should not be compared casually with pass rates from other benches.

Limits

The paper notes that each configuration was evaluated once per task, without variance estimates, which matters for closely ranked systems. In the transfer test, the source-written workflows include worker instructions, which may carry task knowledge as well as organization, so the result does not isolate topology or coordination as the cause. The “first benchmark…” wording is from the official announcement, not a third-party designation. The code repository still had only a handful of stars at announcement time; reproduce from the official page, the paper, and the repo README.

1. Read the official project page https://lab.einsia.ai/workflowarena/ and the announcement https://x.com/EinsiaAI/status/2120162063905681632 ;
2. Check task counts, domain splits, budget, and scoring definitions in arXiv:2611.12856 (https://www.alphaxiv.org/abs/2611.12856) before quoting Overall Resolved or the continuous score ;
3. For reproduction or submissions, use the official code at https://github.com/Einsia/WorkflowArena and record the model, harness (the paper uses Claude Code and Codex), and effort setting ;
4. Do not mix this score with Quant-Arena or other Arena / Bench numbers; figures on this page only repeat the official paper and announcement, so re-check upstream before citing them elsewhere.