Google unveils Budget Tracker and BATS framework to rein in AI agent costs

Google’s AI research division, in collaboration with researchers from New York University and UC Santa Barbara, has launched a new tool and framework that they claim can rein in agent behavior by placing explicit limits on how much compute agents can consume and how freely they can invoke tools.

The tool, called Budget Tracker, is a plug-in module that injects continuous budget awareness into an agent’s reasoning loop, explicitly signaling how much token and tool-call budget remains so the agent can condition its actions on real-time resource availability, the researchers explained in a paper.

How autonomous agents choose to think, loop, and act when completing a task or user request is quietly inflating enterprise AI budgets, which is increasingly turning into a CIO nightmare.

A research report from IDC cited 92% of 318 decision makers surveyed in the US, UK, and Ireland saying that the cost of their deployed AI agents was higher than expected, with inference being the most common cause.

Another report — Greyhound CIO Pulse 2025 — found that 68% of digital leaders surveyed said that they hit major budget overruns during their first few deployments of agents, and nearly half of these leaders pointed to runaway tool loops and recursive logic being the driver behind the overruns.

That budget signal doesn’t stop at individual agents. Google researchers have also launched a framework, named Budget Aware Test-time Scaling (BATS), that builds on the tracker-driven awareness in an agent and tries to adapt the same to a larger multi-agent system.

This enables large, multi-agent systems to determine when it’s worth “digging deeper” into a promising line of reasoning versus when to pivot to alternative paths, thereby improving cost-performance trade-offs under tight budget constraints, the researchers wrote.

Practical path to TCO for agentic AI

Analysts say the tool and framework could offer CIOs a practical path to understand and control costs around agentic deployments.

“The tool and the framework address one of the most pressing questions facing CFOs and CIOs: the true total cost of ownership (TCO) for agentic AI operations. And for enterprise adoption of any technology, TCO is very important,” said Pareekh Jain, principal analyst at Pareekh Consulting.

“Most CIOs would rather accept a predictable agentic system that delivers 90% accuracy at a capped cost than a higher-accuracy system whose per-run cost can vary wildly from a few cents to several dollars,” Jain added.

Seconding Jain, IT management consulting firm Avasant’s research director, Gaurav Dewan, pointed out that many enterprises deploy agents under the assumption that inference is “cheap enough” at scale.

In practice, however, Dewan said, costs grow non-linearly as agents are often deployed without hard budget caps, execution limits, or reasoning constraints.

He further noted that spending is amplified in multi-agent architectures as agents delegate tasks to other agents.

The tool and the framework mark a “big shift” from approaches that similar tools take, according to Greyhound Research Chief Analyst Sanchit Vir Gogia.

“Right now, the tooling landscape is a mess. Plenty of platforms say they optimize agent efficiency, but few actually engage with the problem at runtime. Most solutions in the market operate on a delay; they aggregate usage data, send reports, or shut things down when it’s already too late. That’s not cost governance. That’s damage control,” Gogia said.

The analyst was referring to tools such as LangSmith (LangChain) & Helicone, which offer observability by logging how much an agent spent.

In contrast, Google’s Budget Tracker and BATS frameworks introduce budget awareness directly into the decision loop, giving agents real-time feedback about what they have spent, what’s left, and what’s worth doing next, that too without the need to babysit the agent, Gogia said.

In the absence of a framework like BATS, enterprises typically had to cobble together throttling rules, hard-coded prompt tweaks, and caching strategies to control costs around the agentic system, Gogia added.

Caveats to enterprise adoption

Despite the differentiated approach of the new tool and framework, analysts caution that they might not be a silver bullet in tackling cost overruns around agentic deployments on their own.

CIOs, according to Dewan, are likely to favor the tool or framework if Google integrates natively with existing agent frameworks and offers policy-driven controls spread granularly across all agents, workflows, and business units. “What CIOs need is enforcement, observability, and auditability in one place,” Gogia said. “They want to know what the agent spent, why it made that decision, what path it took, and whether it stayed within bounds. Especially in regulated sectors, where tool usage may touch PII or trigger downstream financial systems, audit trails are non-negotiable.”