Princeton finds agents brittle

Princeton’s CS group posted “Towards a Science of AI Agent Reliability” to arXiv on Feb. 18, 2026, authored by Stephan Rabanser, Sayash Kapoor, Peter Kirgis, Kangheng Liu, Saiteja Utpala and Arvind Narayanan. (arxiv.org) The paper formalizes reliability into four dimensions—consistency, robustness, predictability, and safety—and operationalizes those into twelve concrete, measurable metrics borrowed from aviation, nuclear, and automotive engineering practice. (arxiv.org) The team evaluated 14 agentic models across two complementary benchmarks and reports that recent accuracy/capability gains translated into only modest improvements in the paper’s multi-dimensional reliability score, with overall reliability trending slowly over time. (arxiv.org) Experimental analyses revealed large run-to-run variance, high sensitivity to input perturbations, poor confidence calibration, and agents’ failure to abstain or fail-safe—properties that mean single success-rate metrics can hide catastrophic operational differences. (arxiv.org) Benchmarks used include GAIA (a 450-question general-assistant suite testing web browsing and tool use) and τ-bench (tool-agent-user scenarios with retail and airline domains and a pass^k multi-trial reliability metric), and the study’s evaluations span mainstream vendor model families including offerings from OpenAI, Google, and Anthropic. (hal.cs.princeton.edu) The authors published evaluation code and an interactive reliability dashboard and explicitly recommend adopting safety‑critical practices—fault injection, multi‑trial/pass^k testing, calibration and abstention thresholds, and bounded-harm severity monitoring—for agent development and deployment pipelines. (cs.princeton.edu)