Reproduce “DelegationBench: Measuring When AI Agents Should Ask Before Acting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shiva Pochampally
DelegationBench measures whether AI agents make appropriate permission decisions. It uses matched scenarios and compares judgments of proposed actions with tool use, separating agreement, responsiveness, and asking behavior. Across the same 48 scenarios, all ten models asked less often with tools; Gemini 3.5 Flash-Lite asked in 4.2% of tool runs versus 47.5% of judgments. These results describe benchmark conditions, not deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.