CEO-Bench — Can agents handle 500-day startup runs?
arXiv researchers benchmark language model agents on sustained, multi-skill tasks that mirror real-world complexity.
CEO-Bench drops an agent into a simulated 500-day startup: it must manage pricing, marketing, budgets, and pivots amid uncertainty and incomplete information.
Tests what current LLM agents can't yet do — long horizons, noisy environments, adaptation, and coordinating moving parts toward one goal.