Hand an AI agent a four-hour budget and a workspace full of real professional artifacts, then let it decide: grind through the job alone, or spin up a team of its own clones, farm out the work, and stitch the pieces into one deliverable? That is the exam room WorkflowArena builds, and the best student in the class fully completes only 36.95% of the tasks. The benchmark paper turns "multi-agent orchestration" — a phrase that usually lives in architecture diagrams — into something you can score.
It tests organization, not problem-solving
WorkflowArena packs 145 long-horizon tasks into 8 professional domains — software engineering, enterprise operations, research, formal mathematics, finance, law, content and design, and 3D production — with 17 task forms; the project is funded by Navers Lab and Einsia.AI. Each task provides an initial workspace and demands a deliverable made of interdependent components, graded by a hidden verifier. A submission whose parts are individually correct but never reconciled into a consistent whole does not pass.
The protocol is the interesting part. The model gets a Workflow tool that launches background workflows with up to 16 concurrent workers, and every worker is a fresh instance of the agent itself. Whether to delegate at all, how many workers to launch, and what context each one receives are all decided by the model at run time — and the time spent organizing counts against the same four-hour budget. Working sequentially alone is a perfectly valid workflow. Every deliverable can therefore be read next to the organization that produced it.
The scoreboard: a 36.95% ceiling and heavy specialization
The paper evaluates 9 models in 25 model-effort configurations: Claude Opus 5.5 and Sonnet 5.5 in Claude Code, GPT-6 Astra and GPT-6.1 Sol in Codex, plus GPT-6 Luna, Kimi K3.1 Pro, GLM-5.3 Flash, DeepSeek V4.1 Flash, and Doubao-Seed-Evolving.
The strongest configuration, Opus with Claude Code at max effort, fully resolves 36.95% of tasks on average, and no configuration leads every domain. Formal mathematics comes closest to saturation, with a best resolved rate of 95.83%; the ceilings in law and finance sit at just 9.09% and 18.75%. The gap between partial credit and shippable work is stark: law tasks average a continuous score of 70.7 across configurations while only 2.9% fully pass, and Opus at high effort scores 96.98 in software engineering yet resolves just 17.65% of SWE tasks — models can write a mostly-correct analysis but rarely deliver one that clears every acceptance criterion. Content and 3D deliverables are graded against rubrics by Gemini-3.8-Flash, while the other domains use programmatic verifiers. Chinese models trail visibly on this table: Kimi K3.1 Pro at 16.48%, Doubao at 10.84%, DeepSeek V4.1 Flash at 6.08%, and GLM-5.3 Flash at 4.43%. Effort sensitivity also differs: across five settings Opus swings from 22.02% to 36.95%, while Astra varies only between 27.34% and 30.80%.
Being good at the task is not being good at managing it
The key experiment swaps blueprints: on a challenging subset of 32 tasks, the execution models (Kimi, DeepSeek, Luna, Doubao) are held fixed while only the workflow author changes. Workflows written by Opus improve all four executors, adding 6.41 points on average in the static mode — Kimi climbs from 35.75 to 46.61, DeepSeek from 18.37 to 25.24. Astra's blueprints, by contrast, average −2.53 points in the static mode and drag down both Kimi and DeepSeek, even though Astra on its own (27.34% resolved at max effort) is far stronger than any of these executors. The ability to do the work well does not automatically become the ability to organize others to do it.
The data also pins down two management styles. At high effort, Opus averages 0.86 workflow calls per task with 5.55 workers each — few batches, big teams — while Astra averages 2.09 calls with just 1.72 workers, and about 99% of its workflows contain at most four workers: high frequency, small squads.
So what
This benchmark turns orchestration from marketing language into a measurable, transferable asset: the same weak models gain more than 6 points just by plugging into a better blueprint. The cost math is blunt: Opus at high effort resolves 34.11% at $6.66 per task, while Astra at xhigh spends $29.26 to resolve 30.80% — 4.4 times the cost for 3.31 points less. For anyone building agents, organization is the next battleground beyond single-model benchmarks; for anyone choosing one, ask whether you are hiring a strong worker or a competent manager.