Princeton’s SAgE group published a survey of 33 emerging trends in AI agent evaluation. It divides the trends into methodology, scope, and evaluator roles. The paper argues current benchmarks often fail to capture real-world agent performance. It highlights risks like contamination, gaming, and reward hacking in evaluations. The work notes that evaluator choices shape results as much as the model itself. It introduces concepts like Pareto frontiers for cost versus accuracy. The survey calls for broader participation to improve evaluation rigor. A full run of one benchmark costs about $50,000 in API credits. The group provides interactive tools to explore the trends.

• 33 trends in AI agent evaluation
• 1 full MirrorCode run costs ~$50,000 in API credits