Recently,昆仑万维Skywork and Nanyang Technological University jointly launched an innovative hierarchical multi-agent collaboration framework called AgentOrchestra, which is hailed as the "AI Symphony Orchestra." This framework marks a significant step forward in agent collaboration within the field of general artificial intelligence. Faced with complex tasks in the real world that challenge single models or agents, AgentOrchestra mimics the collaborative model of a symphony orchestra, allowing agents with different expertise to work together to overcome challenges.
Traditionally, general AI models and single-agent systems struggle with complex tasks such as multi-step reasoning and cross-modal information integration. The emergence of AgentOrchestra aims to address these issues. This framework adopts a hierarchical architecture, where the top-level "conductor" agent is responsible for global planning and task decomposition, assigning subtasks to lower-level "musicians" agents with different specialties, such as the Deep Researcher Agent for information retrieval, the Browser Use Agent for web operations, and the Deep Analyzer Agent for deep analysis. These agents collaborate efficiently under the coordination of the "conductor" to complete complex tasks.
AgentOrchestra not only demonstrates advanced design concepts but also showcases outstanding performance in practical applications. The framework supports high scalability, easily handling expansions across different fields and new features; provides a unified tool interface to enable cross-modal understanding and reasoning of multimodal information; adopts a modular design to enhance system scalability and maintainability; and achieves efficient collaboration among agents through hierarchical planning and dynamic role allocation.
In terms of technological innovation, AgentOrchestra also shows remarkable performance. The comprehensive introduction of asynchronous coroutine technology supports high-concurrency collaboration among multiple agents, significantly improving system response speed and throughput. The framework natively supports seamless switching between mainstream commercial large models and open-source local models, balancing capability, privacy, and cost. At the same time, it fully adapts to the Model-Context Protocol, enabling smooth integration of various agents and tools in both local deployment and cloud environments.
In authoritative general intelligent benchmark tasks, AgentOrchestra achieved excellent results, breaking records in multiple core indicators, and its overall performance exceeded several commercial and open-source agent systems. Whether in information retrieval, web operations, or complex reasoning and multimodal analysis, AgentOrchestra has demonstrated strong collaborative efficiency and innovation capabilities.
Currently, the complete code and detailed technical report of AgentOrchestra have been open-sourced for developers and researchers to reference. The Skywork team at昆仑万维 stated that they will continue to improve the system's efficiency and the self-adaptive collaboration capabilities between agents, expand more specialized agents and practical tools, and promote the application of AgentOrchestra in more real-world scenarios such as scientific research, office automation, and knowledge management. This release not only sets a new benchmark for the generality and practicality of AI systems but also paves the way for agent collaboration to enter a new era of general intelligence.
🔗 Paper Link:
https://arxiv.org/pdf/2506.12508
🔗 Project Code:
https://github.com/SkyworkAI/DeepResearchAgent