Multi-agent AI // abstraction for complex tasks
Future of AI Agents is multi-agent // heterogeneous multi-agent systems
Through the collaboration of multiple agents possessing diverse expertise and tools, multi-agent systems achieve impressive progress in solving real-world problems. Given the user queries, the meta-agents, serving as the brain within these systems, are required to decompose the queries into multiple sub-tasks that can be allocated to suitable agents capable of solving them, so-called agent-oriented planning. In this study, we identify three critical design principles of agent-oriented planning, including solvability, completeness, and non-redundancy, to ensure that each sub-task is effectively resolved, leading to satisfactory responses to the original queries. These principles further inspire us to propose a novel framework for agent-oriented planning in multi-agent systems, leveraging a fast task decomposition and allocation process followed by an effective and efficient evaluation via a reward model. During the planning process, the meta-agent is also responsible for evaluating the performance of the expert agents, making timely adjustments to the sub-tasks and scheduling as necessary. Besides, we integrate a feedback loop into the proposed framework to further enhance the effectiveness and robustness of such a problem-solving process. Extensive experiments demonstrate the advancement of the proposed framework in solving real-world problems compared to both single-agent systems and existing planning strategies for multi-agent systems.
Design Principles We further identify three critical principles that guide the design of our framework for effective and efficient agent-oriented planning in multi-agent systems:
• Solvability. Each sub-task qi ∀i ∈ [m] should be independently and completely resolvable by at least one single agent within the multi-agent system, ensuring that the response for each sub-task can be reliable. If a sub-task does not satisfy solvability, the meta-agent is expected to take some modifications or further decomposition.
• Completeness. The array of sub-tasks {q1, …, qm} should include all necessary information from the original user query Q, which ensures that the aggregation of responses of these sub-tasks can effectively yield a comprehensive answer to the user query. If the decomposed sub-tasks fail to satisfy completeness, the decomposition process should be revisited by the meta-agent.
• Non-Redundancy. The array of sub-tasks {q1, …, qm} should not include redundant elements, avoiding those task executions that are either irrelevant to resolving Q, or duplicated. The principle of non-redundancy promotes that the sub-tasks form a minimal effective set necessary to address the user query, enhancing overall efficiency.
In this study, we propose a novel agent-oriented planning framework for multi-agent systems, following three critical design principles to ensure that the meta-agent can effectively decompose the user query into several sub-tasks for producing satisfactory responses. The proposed framework utilizes a fast decomposition and allocation process, which relies on the ability of LLMs to generate an intermediate schedule efficiently. After that, a reward model and the representative work mechanism are employed to evaluate these intermediate results, routing three different paths for executing the sub-task or making necessary modifications to align the sub-task with agents, such as replan the sub-task, plan in detail, and re-describe. Extensive experiments demonstrate that the proposed agent-oriented framework achieves significant improvements over both the existing single-agent and multi-agent baseline methods. We provide discussions on the contributions of different components in the framework and the potential for improvements in agent-oriented planning. We will release the source code to promote further research in the community
