SOURCE-LINKED INTELLIGENCE

Agents Trust Tools Too Much: Measuring Reliance on Unreliable Tools

arXiv · AI, language, vision and robotics · article · Sep 4, 2026 · UTC

Existing evaluations of tool-using agents primarily measure whether an agent can successfully complete diverse tasks with tools. These evaluations generally assume that tools return reliable information. However, tool returns in real-world systems can be plausible yet incorrect. We investigate how agents respond to unreliable tool returns by evaluating fourteen LLMs using three tools-web search, LLM sub-agent delegation, and code execution. For each tool, we corrupt its returns and measure whether agents adopt the corrupted content in their final answers. Agents exhibit high levels of overtrus

Read original source ↗ Open in workspace

recordType

paper

region

Global