Title: BAGEN: Are LLM Agents Budget-Aware?

Abstract

As agents consume increasing amounts of resources, current practices typically measure costs only post-execution. This study argues that a truly Budget-Aware Agent (BAGEN) should utilize the budget as an active control mechanism rather than merely a passive metric of expenditure. We begin by systematically categorizing budget estimation into two types: internal budgets, derived from agent computation, and external budgets, stemming from agent actions. Furthermore, we define budget-awareness through the lens of progressive interval estimation. In this framework, an agent is expected to forecast upper and lower bounds for the remaining budget at every stage of its plan, issuing alerts when success becomes improbable.

Using a rollout-replay protocol for scoring, we identified consistent failure modes across four distinct environments and five leading frontier agents. Our findings reveal four key insights: (1) There is only a weak correlation (r=0.35) between general agent strength and budget-awareness, indicating that powerful agents are not inherently budget-conscious. (2) Frontier models exhibit a persistent tendency toward over-optimism; they continue to expend resources on tasks with low success probabilities rather than notifying users early. (3) The signal for budget awareness is both actionable and trainable. Implementing early stopping on failed trajectories reduces token usage by 28–64%, while Supervised Fine-Tuning (SFT) combined with Reinforcement Learning (RL) enhances both early stopping and alerting behaviors. (4) Achieving precise interval calibration remains difficult, with interval coverage reaching a maximum of 47% even after applying SFT+RL.

Project page: https://ragen-ai.github.io/bagen/

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC