“Performance on METR's agentic tasks using the same GPT-4 base model improved from 5% to nearly 40% through post-training, tool use, and agent scaffolding.”