Meta-Budgeting Agents: Learning to Set Their Own Tool-Call Budgets

Research Question: Can tool-augmented agents learn to predict and dynamically request their own tool-call budgets at inference time, and does this adaptive budgeting improve cost-performance tradeoffs compared to externally imposed budgets?

Hypothesis: Agents empowered to predict and justify their required tool-call budget based on task complexity and their own uncertainty will use resources more efficiently, outperforming both fixed-budget and static budget-aware agents on cost-performance metrics.

Experiment Plan: Extend the Budget Tracker system to allow agents to request a budget at the start of each task, providing a natural language rationale. Train (and/or reinforce) agents to align requested budgets with actual task needs using a dataset of web search tasks varying in complexity. Compare three conditions: (1) fixed budgets, (2) externally adaptive budgets (BATS), and (3) agent-predicted budgets. Measure overall task accuracy, average tool calls, and cost-performance Pareto curves. Analyze rationales for budget requests to identify whether agents develop interpretable heuristics for resource estimation.

References:

• 1. Liu, T., Wang, Z., Miao, J., Hsu, I.-H., Yan, J., Chen, J., Han, R., Xu, F., Chen, Y., Jiang, K., Daruki, S., Liang, Y., Wang, W. Y., Pfister, T., & Lee, C.-Y. (2025). Budget-Aware Tool-Use Enables Effective Agent Scaling.
• 2. Gangopadhyay, B., Wang, Z., Chiappa, A., & Takamatsu, S. (2025). Adaptive Budget Optimization for Multichannel Advertising Using Combinatorial Bandits. Adaptive Agents and Multi-Agent Systems.