Researchers found that handing web search agents more tool calls did not make them smarter, it just wasted the budget faster.
A team at Google Research studied AI agents that search the web using tool calls, and found that simply raising the number of allowed calls hits a performance ceiling because the agents have no sense of how much budget they have left. To fix that, the researchers built the Budget Tracker, a lightweight plug-in that gives an agent continuous awareness of its remaining token and tool-call budget. They then built BATS, or Budget-Aware Test-time Scaling, which uses that awareness to adjust its planning and verification strategy on the fly. The team also formalized a unified cost metric combining token and tool consumption, so different agents can be compared on equal footing.
This lands the same lesson the field learned about raw parameter counts onto a new axis: throwing more actions at an agent, like throwing more parameters at a model, stops paying off once the system cannot tell how much runway it has left. It is a useful corrective for the current habit of measuring agent progress by tool-call counts alone, since the real gains come from budget-aware planning, not budget size. For teams building production agents that pay per API call, that is the difference between a system that improves with more budget and one that just gets more expensive.
Worth noting: this paper is a replacement of an earlier arXiv submission, not a first release, a small reminder that even research on efficient scaling took more than one draft to get right.