Can Budget‑Aware Tool Use Unlock Scalable AI Agents? A Deep Dive

Background and Motivation

In 2025 the large‑language‑model (LLM) community highlighted two major trends: test‑time scaling, which improves performance by “thinking more” or “trying more” rather than increasing parameters, and the rise of agents that can iteratively reason in an environment. A key open question is whether adding more agents always yields better results.

Google’s Two Papers

Google recently released two papers that turn agent scaling into a measurable scientific problem:

Budget‑Aware Tool‑Use Enables Effective Agent Scaling – investigates how to make agents spend less budget while achieving correct outcomes.

Towards a Science of Scaling Agent Systems – asks whether the optimal number of agents and coordination structure can be predicted in advance.

Budget‑Aware Tool Use

Core Pain Points

Simply increasing the budget does not improve performance; agents quickly hit a ceiling without budget awareness.

Tool calls have an economic cost distinct from token usage; a unified metric is needed.

Solution 1: Budget Tracker (Plug‑and‑Play)

The Budget Tracker plugin writes the remaining/used budget into the prompt each round, requiring zero additional training. It automatically switches between “broad search” and “precise strike” strategies based on the budget level.

Results (using BrowseComp with Gemini‑2.5‑Pro):

Scaling from budget 10 to 100 continues to improve performance, whereas the baseline without a tracker saturates at 100.

At equal accuracy, cost drops by 31 % (search cost ↓ 40 %, browsing cost ↓ 21 %).

Solution 2: BATS Framework (Budget‑Aware Test‑time Scaling)

BATS adds two modules:

Planning : writes “remaining tool calls” into a checklist to decide whether to dig deeper or change direction.

Self‑Check : after producing an answer, uses the remaining budget for reverse verification; failures are compressed into memory and a new path is opened.

Experimental outcomes on three information‑retrieval benchmarks show BATS consistently outperforms both parallel and serial scaling approaches while incurring lower actual cost.

Scaling Science: The Break‑Even Point of Multiple Agents

The authors evaluated 180 configurations covering four real‑world agentic benchmarks (finance, web, Minecraft planning, office workflow), nine LLMs from three families, and four multi‑agent system (MAS) architectures (Independent, Centralized, Decentralized, Hybrid). All configurations matched token budgets to isolate scaling effects.

Key Empirical Laws

Tool‑Coordination Trade‑off : β = –0.267 (p < 0.001). When more than eight tools are used, MAS overhead grows exponentially – use sparingly.

Ability Saturation Point : Adding agents beyond the point where a single agent contributes > 45 % of total performance yields negative returns; strengthen the single agent first.

Error Amplification : Independent architectures amplify errors by 17.2×, whereas Centralized reduces amplification to 4.4× – avoid “bare parallel” without verification.

Quantitative Prediction Model

A mixed‑effects model built on 20 observable features (tool count, single‑agent baseline, efficiency, redundancy, error amplification, etc.) predicts the optimal architecture with cross‑validated R² = 0.524 and MAE = 0.089. It correctly identifies the best architecture for 87 % of held‑out configurations.

The proposed online calculator takes task complexity T, single‑agent baseline PSA, and model Intelligence Index as inputs and outputs the architecture expected to deliver the highest performance.

Conclusion

The two papers transform agent scaling from intuition‑driven practice into a predictable, metric‑based science. Budget‑aware tool use and the BATS framework demonstrate that thoughtful budget management and self‑verification can substantially improve performance and reduce cost, while the scaling laws and predictive model provide actionable guidance for designing multi‑agent systems.

https://arxiv.org/pdf/2511.17006
Budget-Aware Tool-Use Enables Effective Agent Scaling

https://arxiv.org/pdf/2512.08296
Towards a Science of Scaling Agent Systems