Agent Initiative Benchmark: Evaluating Proactive AI Models in a Synthetic Workplace

Deepnote researchers have published a benchmark comparing 12 AI models on their ability to act as proactive workspace agents within a synthetic company environment named PageMantle. The evaluation reveals that OpenAI models lead on overall score and cost efficiency, Anthropic models excel in speed, and the primary weakness across nearly all models is knowing when to stay silent rather than intervening unnecessarily. Open article →