Einsia Launches WorkflowArena to Test How Well AI Models Orchestrate Teams of Agents
Agent harnesses now let a model spawn its own workers and write the workflow that coordinates them. But can it actually orchestrate them well? Introducing Einsia AI's WorkflowArena, the first benchmark for autonomous multi-agent workflow orchestration on real, long-horizon work. - 145 tasks across 8 professional domains, from SWE and Lean proofs to law and 3D - 25 configurations of 9 frontier models in Claude Code & Codex - Every delegation recorded: we score the deliverable and see the workflow behind it 馃寪 Project: t.co/j98U5JTv2n 馃捇 GitHub: t.co/oQg0e5jyyu 馃搫 Paper: t.co/MauLtxSuta 猬囷笍 Read the following threads to go through the key findings
Context
Agent tools now let one model spawn and direct other agents, and this benchmark tries to measure whether models are actually good at managing that work.
Coding agents such as Claude Code and Codex can now launch helper agents and write the workflow that coordinates them. Most benchmarks score a single model on a single task, not how well it plans and delegates work across many agents. Einsia AI is the team behind the new benchmark.
Deep dive
Einsia AI introduced WorkflowArena, which it calls the first benchmark for autonomous orchestration of workflows with many agents on real, long horizon work. The question it sets out to answer is whether a model that can spawn its own workers and write the workflow that coordinates them can actually orchestrate them well.
According to the post, the benchmark has 145 tasks across 8 professional domains, ranging from software engineering and Lean proofs to law and 3D work. The overview image lists domains such as operations, science, math, finance, law and content, with example tasks including a Blender scene, a short film, a Lean 4 proof, a weekly payment run and a peer analysis of company filings.
Einsia says it tested 25 configurations of 9 frontier models running inside Claude Code and Codex. Every delegation is recorded, so the team scores the final deliverable and can also inspect the workflow behind it. The project page, code on GitHub and a paper on arXiv are linked from the post, and Einsia says its key findings are in a follow up thread.
These results come from the benchmark's own authors and have not been independently reproduced.
What to expect next
Watch for model makers and agent tool builders to cite WorkflowArena scores, and for outside teams to run their own systems on the public code.
The open question is whether orchestration skill tracks raw model strength or depends mostly on the harness. If the benchmark catches on, expect labs to start tuning models specifically for planning and delegating work across agents.