Agent trace storage and replay

Warp replays your stored agent runs to decide which model gets the next ticket

The replay only works on runs Warp's infrastructure already recorded, and correctness is graded by a judge model against a rubric you write. Warp's own 30-task bake-off cost $2,130.57. It finished in under four hours.