Sayash Kapoor

Sayash Kapoor studies a foundational weakness in the market for AI agents: impressive demonstrations often reveal little about reliability, cost or real-world usefulness. His recent work builds evaluation systems intended to replace isolated scores with reproducible evidence across tasks, models and repeated runs.

Sayash Kapoor’s work begins where many AI demonstrations end. An agent completes a task once; the result circulates as evidence of capability; and questions of reliability, cost and reproducibility arrive later. As a Princeton computer-science doctoral candidate, Kapoor has helped build a more demanding science of agent evaluation. In 2025, AI Agents That Matter, co-led with Benedikt Stroebl, argued for benchmarks that account for deployment costs, overfitting and meaningful downstream usefulness. In 2026, Kapoor and Stroebl were co-first authors of the Holistic Agent Leaderboard, an open evaluation platform designed to compare agents across multiple benchmarks while exposing cost, variance and methodological choices. Kapoor also co-authored work on agent reliability that treats performance as a distribution across repeated trials rather than a single pass rate. Alongside this technical programme, he contributed to the 2025 and 2026 International AI Safety Reports and research on independent flaw disclosure. His contribution is deliberately corrective: not a claim that agents are unimportant, but a demand that the evidence used to judge them become as sophisticated as the systems being sold.