Agents solve tasks by repeatedly calling a large language model with growing histories of reasoning, actions, and tool results. Each call is slow because output comes one small piece at a time. uses a smaller model to guess pieces and the main model to check them. Overall averages hide which phases are slow.
TraceSpec records decision histories and replays saved states to locate where drafts fail. It trains the draft model on intermediate steps, not just final answers, and devotes a fixed training budget to the states where rejection is most likely.
This raised replay throughput by 12.5% to 22.6% over base. Faster output from the same target model makes agents cheaper to run with no behavior change.