TLDR: HAWK is a new hierarchical framework for multi-agent AI systems, addressing challenges in cross-platform interoperability, dynamic task scheduling, and efficient resource sharing. It comprises five layers (User, Workflow, Operator, Agent, Resource) and sixteen standardized interfaces, ensuring modularity, extensibility, and adaptive scheduling. Demonstrated by CreAgentive, a novel-generating prototype, HAWK significantly improves task throughput and system controllability. Future applications are envisioned across healthcare, government, finance, and education.
In the rapidly evolving world of artificial intelligence, multi-agent systems are becoming increasingly common, powering everything from intelligent customer service to autonomous vehicles. However, these systems often face significant hurdles, including difficulties in getting different AI agents to work together across various platforms, managing tasks efficiently, and sharing resources effectively. These challenges stem from a lack of standardized interfaces, rigid collaboration frameworks, static scheduling policies, and insufficient ways for agents to synchronize their information.
To tackle these complex issues, researchers have introduced HAWK, which stands for Hierarchical Agent Workflow. HAWK is a new, modular framework designed to streamline how multiple AI agents collaborate. It offers a comprehensive, end-to-end solution that covers everything from understanding user requests to orchestrating workflows, intelligently scheduling tasks, invoking necessary resources, and ensuring data is synchronized between agents.
HAWK is structured into five distinct layers, each with specific responsibilities, ensuring a clear separation of concerns and promoting modularity. These layers are:
User Layer
This is the entry point for users, providing interfaces to submit task requests. It translates natural language or graphical inputs into structured descriptions that the system can understand and process.
Workflow Layer
At the heart of HAWK, this layer is responsible for planning, executing, monitoring, and optimizing agent workflows. It includes an adaptive scheduling and optimization module that uses real-time feedback to adjust strategies, maximizing resource utilization and overall efficiency. This layer ensures that workflow models are selected appropriately and directs the execution of decomposed tasks to the layer below.
Operator Layer
This layer manages the scheduling and execution of specific tasks within the agent workflow. It comprises six core modules: Environment (for contextual configurations), Memory (for state maintenance), Task Management (for distribution and monitoring), Task Optimizer (for dynamic strategy adjustment), Reasoning (for complex reasoning), and Security (for privilege control and data protection). It acts as a bridge between scheduling logic and execution.
Agent Layer
This layer focuses on managing agents themselves, handling their specification, publication, registration, and discovery. By decoupling agent management from the operational logic, HAWK allows agents to be deployed flexibly across diverse environments, enhancing system robustness and fault tolerance.
Also Read:
- Mod-X: Building Decentralized Communication for AI Agent Ecosystems
- The Rise of AI Agents and Agentic AI: A New Era for Manufacturing
Resource Layer
Serving as the foundational support, this layer provides a unified abstraction over various resources, including heterogeneous data sources (structured and unstructured), large models (language and image models), physical devices (cameras, robotic arms), and third-party services and tools. It simplifies resource access for the Operator and Agent layers, boosting compatibility and scalability.
HAWK also defines sixteen standardized interfaces that ensure seamless interaction and interoperability between its various modules and layers. These interfaces facilitate end-to-end coordination, from initial user requests to autonomous agent collaboration and resource invocation.
To demonstrate HAWK’s capabilities, the researchers developed CreAgentive, a prototype multi-agent system for generating novels. CreAgentive showcases how HAWK can orchestrate specialized agents to collaboratively write a novel, managing long-term story goals and short-term chapter objectives. The system involves an Environment Agent to maintain story state, a Decision Agent to select the best narrative trajectories, a Writer Agent to generate chapters using large language models (LLMs), and an Ending Determination Agent to check if the story is complete.
Initial evaluations with CreAgentive have shown promising results, indicating HAWK’s effectiveness in improving task throughput, simplifying resource invocation, and enhancing system control. The study also explored the performance of different LLMs (Deepseek-V3, Qwen/QwQ-32B, and GLM-4-9B) in maintaining multi-chapter continuity, suggesting that a hybrid approach using the optimal model for each specific module could yield the best overall performance. While the prototype faced challenges like LLM hallucinations and occasional rule violations in generated content, the modular design of HAWK allows for future enhancements to address these issues.
Looking ahead, HAWK is envisioned for broad applications across various sectors, including healthcare (for diagnostic systems and clinical workflow coordination), government services (for smart cities and public safety), economics and finance (for analysis and forecasting), and education (for personalized learning platforms). Its robust architecture is designed to orchestrate millions of diverse agents in distributed environments, laying the groundwork for a unified, collaborative AI infrastructure. For more details, you can read the full research paper here.
