Budget-Aware Multi-Agent Collaboration: Dynamic Resource Sharing Among Agent Teams

Research Question: Can a team of tool-augmented agents collaboratively allocate and trade tool-call budgets to maximize collective task performance across a diverse set of tasks?

Hypothesis: Multi-agent systems with budget-trading mechanisms (e.g., auctions or reinforcement learning–based negotiation) will outperform independent agents and static budget partitions, achieving higher overall accuracy and more efficient budget utilization.

Experiment Plan: Design a simulation where several agents, each with different strengths (e.g., web search, code, summarization), are given a shared tool-call budget. Implement budget-sharing protocols: auctions, negotiation, or learned policies via multi-agent reinforcement learning. Tasks are drawn from complex benchmarks like MCPVerse or UrbanMUDA, requiring different combinations of expertise. Measure collective task performance, individual agent contributions, and emergent budget-sharing strategies. Analyze whether certain agent roles or specializations lead to consistent budget imbalances or cooperation patterns.

References:

• 1. Liu, T., Wang, Z., Miao, J., et al. (2025). Budget-Aware Tool-Use Enables Effective Agent Scaling.
• 2. Peng, B., Wang, Y., Feng, C., Xia, X., & Li, P. (2025). UrbanMUDA: an LLM Agent-based Site Selection Approach for Urban Military Unit Deployment.
• 3. Mohanty, A., Nayak, S., Yang, T., Rathore, R. S., Mo, D., & Wang, L. (2024). A Hybrid Reinforcement Learning Framework for Dynamic Resource Allocation in Malware Analysis Systems.