Towards a Science of AI Agent Reliability
· Revised 2 June 2026; accepted at ICML 2026
Summary
The authors evaluate 15 models on two agent benchmarks using 12 metrics spanning consistency, robustness, predictability, and safety. Higher average capability brings only small gains on several reliability dimensions, so benchmark progress need not translate proportionally into dependable autonomy.
Why it matters
The paper turns reliability into a multidimensional economic deployment question: firms need repeated, predictable performance and low supervision, not only a higher average task-success rate.
What to keep in mind
The results are specific to the selected benchmarks and scaffolds rather than production incidents. The reliability threshold required for deployment also varies by task, error cost, and available human oversight.