Budget-Aware Tool-Use Enables Effective Agent Scaling (2511.17006v1)

Abstract: Scaling test-time computation improves performance across different tasks on LLMs, which has also been extended to tool-augmented agents. For these agents, scaling involves not only "thinking" in tokens but also "acting" via tool calls. The number of tool calls directly bounds the agent's interaction with the external environment. However, we find that simply granting agents a larger tool-call budget fails to improve performance, as they lack "budget awareness" and quickly hit a performance ceiling. To address this, we study how to scale such agents effectively under explicit tool-call budgets, focusing on web search agents. We first introduce the Budget Tracker, a lightweight plug-in that provides the agent with continuous budget awareness, enabling simple yet effective scaling. We further develop BATS (Budget Aware Test-time Scaling), an advanced framework that leverages this awareness to dynamically adapt its planning and verification strategy, deciding whether to "dig deeper" on a promising lead or "pivot" to new paths based on remaining resources. To analyze cost-performance scaling in a controlled manner, we formalize a unified cost metric that jointly accounts for token and tool consumption. We provide the first systematic study on budget-constrained agents, showing that budget-aware methods produce more favorable scaling curves and push the cost-performance Pareto frontier. Our work offers empirical insights toward a more transparent and principled understanding of scaling in tool-augmented agents.

• The paper introduces a Budget Tracker plug-in and BATS framework to integrate explicit budget awareness, enabling agents to adaptively manage tool call costs and scale effectively.
• The methodology employs a unified cost metric for token and tool-call expenses, optimizing exploration strategies and demonstrating significant accuracy improvements.
• Empirical results on benchmarks like BrowseComp and HLE-Search show that budget-aware agents outperform traditional baselines through dynamic resource reallocation under varying constraints.

Budget-Aware Tool-Use for Effective Agent Scaling

Motivation and Problem Formulation

Recent developments in LLM-augmented agents have demonstrated strong empirical improvements from scaling test-time compute, yet agentic settings introduce unique complexities beyond traditional chain-of-thought or reflection-based scaling. Tool-using agents such as web search or browsing models must manage not only token-based cognitive effort but also external tool calls, the latter of which incur substantial costs and define the agent's effective information boundary. The central observation of "Budget-Aware Tool-Use Enables Effective Agent Scaling" (2511.17006) is that simply increasing tool-call budgets for agents does not yield consistent or meaningful gains due to a lack of budget awareness. Agents saturate their cost-performance curves early and are unable to effectively utilize additional allocated resources unless made explicitly aware of these constraints.

The paper formalizes agent test-time scaling as a constrained optimization—maximizing answer accuracy under per-tool call budgets, with joint consideration of internal compute (tokens) and external calls—operationalized via a unified cost metric mapping both to their economic costs. The focus is on practical search agents using both search and browse tools, evaluated on standardized information-seeking benchmarks.

Budget Tracker: Enabling Explicit Budget Awareness

To counteract the observed plateau in naive scaling, the authors introduce Budget Tracker, a lightweight, architecture-agnostic plug-in that injects explicit budget state into each agent decision step. Every agent reasoning cycle is augmented with a block specifying the current and remaining budget per tool (Figure 1). This is prompt-only and requires no additional model training or finetuning.

Figure 1: At each interaction round, the agent is provided with its current and remaining budget through the budget tracker before generating the next thinking step and the tool call actions.

By surfacing budget signals continuously, the agent is incentivized to adapt its tool-use behavior in response to resource availability, e.g., utilizing parallel search paths under high budget conditions or prioritizing critical queries when resources are low. Coupled with sequential and parallel test-time scaling paradigms, Budget Tracker consistently enables more effective scaling than non-aware baselines, leading to improvements in accuracy and resource utilization as evidenced by comprehensive ablation and scaling studies (see Figure 2 and Figure 3).

Figure 2: Comparison of Budget Tracker and ReAct in sequential scaling using Gemini-2.5-Pro. With explicit budget awareness, Budget Tracker consistently improves upon ReAct at equal budgets. ReAct plateaus early as it cannot utilize extra resources, while Budget Tracker adapts its spending to gain further improvements and extend the cost-performance frontier.

The adaptive behavior is illustrated in case analyses: under high budgets, agents expand search breadth, while under constrained budgets, they pursue efficient, highly targeted exploration. This directly extends the Pareto frontier of cost-performance, enabling efficiency gains and avoiding premature agent stalling.

Figure 3: Comparison of Budget Tracker and ReAct in parallel scaling using Gemini-2.5-Pro. The left subfigure shows accuracy scaling with increasing parallel runs, while the right subfigure illustrates the corresponding cost–performance trend.

BATS: Budget-Aware Test-time Scaling Framework

While Budget Tracker enables basic resource-aware behaviors, fundamentally optimal scaling requires dynamic orchestration beyond isolated prompt cues. The paper proposes BATS (Budget-Aware Test-time Scaling), an agent framework that systematically internalizes both planning and self-verification with persistent awareness of remaining budgets and accumulated costs.

The orchestration is modular:

• Planning Module: Generates and continually updates a tree-structured, budget-annotated checklist of exploration and verification subtasks. Steps are dynamically marked (pending, done, failed, partial), with allocation strategies adjusting search breadth and verification depth according to available resources.
• Self-Verification Module: After each candidate answer, performs constraint-level backward checks, diagnoses whether constraints are satisfied, unverifiable, or contradicted, and uses budget and trajectory information to decide among SUCCESS, CONTINUE (dig deeper), or PIVOT (switch to alternatives). Summaries and optimization advice are produced to compress context and prevent redundant tool usage.

  Figure 4: Overview of the BATS framework. Given a question and per-tool budgets, the agent begins with budget-aware thinking and planning, structured as a checklist. The agent keeps iterating by reasoning over new information and updated budgets. When an answer is proposed, BATS performs verification and decides to either continue, pivot, or initiate a new attempt with the remaining budget. BATS terminates when any of the budgets are exhausted.

This tightly integrated awareness enables fine-grained, resource-efficient reasoning—e.g., canceling unproductive lines of investigation and reallocating effort to promising alternatives—thereby maximizing accuracy for a given total budget (Figure 5).

Figure 5: Scaling behaviors along (a) total number of tool calls and (b) average unified cost, evaluated on a 200-example subset of BrowseComp using Gemini-2.5-Pro.

Empirical Results and Insights

Comprehensive benchmarking is performed on BrowseComp, BrowseComp-ZH, and HLE-Search. Budget-aware methods are exhaustively compared to strong agentic baselines (e.g., ReAct, self-consistency, state-of-the-art parallel and sequential scaling, and fine-tuned specialist agents), under identical per-tool budgets and using the unified cost metric for analysis. Notably, the BATS framework is entirely training-free yet consistently outperforms strong baselines with significant margins:

• On BrowseComp with 100 tool calls: Gemini-2.5-Pro with BATS achieves 24.6% accuracy, compared to 12.6% for ReAct and 15.7% for training-based WebExplorer.
• On BrowseComp-ZH: 46.0% (BATS) vs. 31.5% (ReAct).
• On HLE-Search: 27.0% (BATS) vs. 20.5% (ReAct).

BATS maintains steeper scaling curves and higher cost-efficiency at all operating points, demonstrating resilience to early stopping and efficient budget utilization without forced exhaustion of allocated resources (Figure 6).

Figure 6: Performance and resource utilization under early stopping on BrowseComp-ZH using Gemini-2.5-Pro. The bars represent the average number of Search and Browse tool calls (left axis), while the line plots indicate accuracy (right axis). While ReAct plateaus and fails to utilize additional budget, BATS effectively scales up tool usage to achieve higher accuracy as the budget increases.

Ablation studies reveal complementary contributions: removing structured planning or self-verification notably degrades performance, underscoring the necessity of tightly coupled, budget-conditioned orchestration.

Case studies further exemplify the qualitative advantages: budget-aware agents dynamically select exploration/verification strategies and adaptively reallocate effort based on observed failure modes and residual resources (Figures 9, 10, 11).

Figure 7: Adaption to high budget constraints. Being aware of the high budget, the agent chooses to expand its search queries starting from a specific keyword, whereas ReAct immediately applies narrow conditions and fails.

Figure 8: Adaption to low budget constraints. Under a low budget, ReAct fails due to an exhaustive search strategy. In contrast, budget tracker helps the agent adapt by prioritizing query efficiency and precision, successfully solving the task.

Figure 9: In BATS, the verification process dynamically adapts to budget constraints. It successfully identifies the cause of failure and provides strategic recommendations based on the current trajectory. Following this advice allows the agents to better utilize the budget and arrive at the correct answer in the next attempt.

Theoretical Implications and Future Directions

The paper makes explicit the gap between naive scaling (more tokens or calls) and cost-optimal scaling in agentic LLMs, establishing budget awareness as a principal factor for efficient search and reasoning under real-world constraints. The unified cost-performance analysis and empirical push on the Pareto frontier have direct implications for deployable AI systems, especially where cost, latency, and external API rate limits are critical.

The approach motivates several directions for theory and systems:

• Joint optimization over multiple interacting constraints (token, FLOPs, tool calls, latency)—extending beyond single-resource awareness toward multi-dimensional adaptive agentic control.
• Automated resource allocation and estimation: predictive strategies for future tool use based on observed problem hardness and context trajectory, integrating explicit utility estimation.
• More sophisticated context and memory engineering for efficient long-horizon reasoning.
• Integration with RL-based and meta-learning approaches for dynamic adaptation to budget and problem characteristics.

Conclusion

Budget awareness is a necessary condition for effective agent test-time scaling in tool-augmented settings. The combination of explicit budget signal injection (Budget Tracker) and holistic orchestration (BATS) enables agents to break cost-performance ceilings, scale adaptively with resources, and maximize utility under real-world deployment constraints. The paper lays foundational methodology for principled, transparent, and efficient agent design and benchmarking.