We ran 153 autonomous runs across 18 frontier models on the nanoGPT optimizer speedrun.
Open Traces to explore 41 curated full agent trajectories, including tool calls, subagents, and scratchpads.