What's this about?
Business Arena is a benchmark that tests LLM agents by making them run a business: infer opportunities from partial signals, commit capital under uncertainty, and adapt to delayed outcomes in a changing market.
What exactly is an "agent benchmark"?
A benchmark is a standardized test for comparing models. An *agent* benchmark tests not just answering questions but taking a sequence of actions toward a goal in an environment. Business Arena's environment is a simulated marketplace, so it measures judgment under uncertainty and long-horizon planning — skills closer to real work than trivia.
Why it matters
As agents move toward autonomous economic tasks, benchmarks that capture messy, delayed, high-stakes decision-making — not just clean puzzles — are how we'll tell which models can actually be trusted to act.