AI agents handle workplace tasks like replying to emails, but first must spot those tasks in a flood of messages. Existing tests skip this.
AutoBuddy-Bench simulates an office with consistent emails, chats, and files. It generates tasks from events like a report request. The AutoBuddy agent was tested with many large language models, AI systems that generate text and make decisions.
Scores ranged from 0.23 to 0.70. Weaker models failed to find tasks; intermediate models found tasks but stopped short; the strongest failed during execution, especially on business-system operations. So discovery and execution must be evaluated separately.