We contributed a bunch of evaluation environments to terminal-bench-3 and I’d like to talk about the workflow of creating complex environments, the difficulties along the way (frontier models are just too good - but also not goot enough), and how terminal-bench-3 approaches this crux.
