EdgeBench: Benchmark for agents working 12-72 hours on tasks

Original: such a beautiful graph of agents learning through environment interactions

Source: x.com ↗