Tokens don’t matter if you have any slow deterministic processes. The bottleneck is already on testing infrastructure in the work I do. Screenshot tests are slow. Enormous codebases are slow. We have whole teams devoted to speeding these things up, so all I can do is get the agent to call the minimum set of commands necessary to check their latest iteration. This of course flies right in the face of the mountains of slow commands the harness tells it to use, resulting in chaos.