Preprint tests speculative execution for tool-using agents

A preprint describes Speculative Macro Commit, a runtime method in which a smaller model predicts and executes future tool-action chains on an isolated environment snapshot while an authoritative model produces the official trajectory. When their next actions match, the system commits the pre-executed steps and observations. Using Qwen3.5 models, the authors report matching sequential-agent accuracy while reducing latency by 18.59% versus sequential execution on a telecom subset and cutting AppWorld wall time by 44.9%, with a small drop in task completion.