Cost-aware stopping cuts agent tool exposure 37%

CAM-DF cut tool exposure to AI agents by 37%, holding task success steady, per a July 2026 arXiv paper. The test spanned five domains and 1,343 tasks.

That's a real budget line for anyone running an agent against a growing MCP tool list. Every extra tool description in a prompt eats context and adds cost before the model does anything useful.

Stopping instead of ranking

The method treats tool acquisition as a stopping problem instead of a ranking problem. It trains offline on the gap between stopping now and continuing to the best next tool call. The sign of that gap labels the decision to stop or continue. The size of the gap weights each error by how much it costs, so bigger mistakes matter more.

It sits in front of an existing tool ranker as a pre-execution plugin. Getting the 37% cut required no fine-tuning of that ranker, per the paper.

A third of the tools most agents load into context are dead weight nobody's auditing. Cost-aware stopping proves it without touching the ranker underneath. The paper doesn't say how the method holds up on tool lists bigger than the ones tested. It also doesn't say whether the stopping signal holds up if the underlying ranker is weaker than the one tested here.

The paper landed on arXiv in July 2026, as MCP tool lists on many agent setups quietly outgrow their context budgets.