MAS-Orchestra
A reinforcement-learning training framework that composes an entire multi-agent system at each step rather than building it incrementally, paired with MASBench, a controlled benchmark for when multi-agent systems beat single-agent ones.
MAS-Orchestra trains an orchestrator with a function-calling abstraction to compose a whole multi-agent system holistically at each step, rather than assembling it piece by piece, and the project reports that it outperforms GPT-5 and Claude Sonnet 4.5 by up to 23 percent across five benchmarks with roughly a tenfold efficiency gain. Its companion MASBench is a controlled benchmark suite built specifically to study when a multi-agent system actually outperforms a single agent, rather than assuming more agents are always better.
The project provides downloadable orchestrator checkpoints for specific benchmark tasks such as AIME math problems, HotpotQA and BrowseComp+, and a live demo for trying its vibe-coding orchestration interactively. It was accepted at ICML 2026.