Two new benchmarks show AI agents fail to improve over long horizons

FM-Bench and E-Commerce Bench, two recently published long-horizon agent benchmarks, both find that current AI models do not learn to improve their strategies over extended simulations and that no single model dominates. FM-Bench has 15 frontier models manage a football club across 20 simulated years with 340 to 400 decision points; on one seed, year-5 rankings correlated only 0.19 with final rankings, and DeepSeek-V4-Pro led at years 5 and 10 but finished 12th overall. E-Commerce Bench, introduced by Qwen, has agents start with ¥100,000 to run online stores for 365 simulated days handling sourcing, negotiation, pricing, and inventory; almost no model tested learns to buy cheaper or improve its strategies over the full simulated year.