OrgAgent: Organize Your Multi-Agent System like a Company

A research team from Tsinghua University and IBM has published a paper on arXiv introducing OrgAgent, a novel framework for organizing multi-agent AI systems. The core innovation is structuring AI agents into a three-tiered hierarchy inspired by corporate organizations: a governance layer for high-level planning and resource allocation, an execution layer where specialized agents solve tasks, and a compliance layer that reviews and controls the final output. This structure is designed to address the open question of how to effectively coordinate multiple large language model (LLM) agents for complex reasoning tasks.

The team rigorously evaluated OrgAgent across various reasoning tasks, LLMs, and execution policies. The results were striking. For instance, when using the GPT-OSS-120B model on the SQuAD 2.0 question-answering benchmark, the hierarchical organization improved performance by over 102% compared to a flat, non-hierarchical multi-agent system. Simultaneously, it reduced token consumption—a key cost and efficiency metric—by nearly 75%. Further analysis revealed that this hierarchical coordination excels in tasks that benefit from stable skill assignment among agents, controlled information flow, and layered verification processes.

Overall, the findings position organizational structure as a critical, previously underexplored factor in multi-agent AI. OrgAgent demonstrates that how agents are organized shapes not just their effectiveness and operational cost, but also their fundamental coordination behavior. This work provides a concrete, high-performing blueprint for developers building sophisticated AI agent systems that need to be both capable and cost-efficient.