Budget-Constrained Abstention and Deferral for Tool-Using LLM Agents

This record archives the SIT320 Advanced Algorithms High Distinction task (Option B) at Deakin University. It contains a research proposal on budget-constrained abstention and deferral for tool-using LLM agents, a comparison of the proposed methods with ReDAct and other related work, and the Jupyter notebook with the simulation. The proposal formulates the act, defer or clarify decision as a finite-horizon constrained Markov decision process with the remaining budget in the state, and proposes three solution methods: an exact oracle, a budget-conditioned learned policy with action masking, and an interpretable threshold schedule. In a fully specified simulation the budget-aware policy never violates the cap and outperforms a ReDAct-style fixed threshold at every budget.