Budget-Constrained Agentic Large Language Models: Intention-Based Planning for Costly Tool Use

TL;DR

This work addresses budget-constrained tool use by modeling it as a sequential decision problem for agentic LLMs and introducing INTENT, an intention-based inference-time planning framework. INTENT combines a lightweight language world model with an ideal-trajectory simulation and a risk-calibrated cost estimator to enforce hard budget constraints without retraining or environment interaction. Empirical evaluation on StableToolBench shows INTENT delivers higher budget-feasible task success than soft-prompt baselines and hard-enforcement methods, with robust performance under dynamic tool prices and tool availability. The results demonstrate that intention-level planning can reconcile strong agentic capabilities with realistic monetary constraints, enabling scalable, budget-aware tool use in non-stationary marketplaces.

Abstract

We study budget-constrained tool-augmented agents, where a large language model must solve multi-step tasks by invoking external tools under a strict monetary budget. We formalize this setting as sequential decision making in context space with priced and stochastic tool executions, making direct planning intractable due to massive state-action spaces, high variance of outcomes and prohibitive exploration cost. To address these challenges, we propose INTENT, an inference-time planning framework that leverages an intention-aware hierarchical world model to anticipate future tool usage, risk-calibrated cost, and guide decisions online. Across cost-augmented StableToolBench, INTENT strictly enforces hard budget feasibility while substantially improving task success over baselines, and remains robust under dynamic market shifts such as tool price changes and varying budgets.

Figures (7)

• Figure 1: Budget awareness of agentic language models on tool cost-augmented StableToolBench. Standalone agents frequently violate hard budget constraints, and prompt-based cost feedback remains insufficient to guarantee budget feasibility or approach the achievable performance upper bound. Our lightweight online planning framework INTENT helps bridge this gap.
• Figure 2: Inference-time planning paradigms for budget-aware agentic tool use. (a) MCTS explores a large stochastic search tree with prohibitive cost. (b) MCO enforces budgets via a single stochastic rollout using a language world model. (c) INTENT extracts the agent’s latent plan through ideal trajectory simulation and applies intention-aware, risk-adjusted cost estimation for stable budget control.
• Figure 3: Performance under varying amounts of oracle training data, simulating the introduction of new tools. Data points correspond to fractions of the full interaction log set (from $1/32$ to $1$). INTENT shows a clear log-linear scaling trend and strong performance even in the low-data regime, across both backbones.
• Figure 4: Robustness to relative price changes of reference tools. We uniformly increase or decrease the prices of reference tools by fixed ratios (from a $50\%$ discount to a $50\%$ markup), while keeping other tools unchanged. INTENT is substantially less sensitive to price perturbations than Prompt across both backbones.
• Figure 5: Performance under varying budget levels. Budgets are scaled by fixed ratios relative to the default setting. INTENT scales effectively with increased budget and achieves competitive performance under tight budgets, across both backbones.