Speculative Macro Commit Pre-Executes Multi-Action Chains and Cuts Agent Wall Time 44.9% on AppWorld

arXiv 2609.03236·medium signalXBlueskyLinkedInCopy link

Tool-using agents lose wall-clock time to serial action-observation turns, not just inference. SMC runs a two-tier system where a large authoritative actor produces the official trajectory while a faster drafter continuously predicts and executes future action chains on an isolated environment snapshot, matching against a macro library of recurring multi-action skeletons mined from training traces; when the actor's next tool call matches the first drafted action, SMC commits the remaining pre-executed steps and their observations. With Qwen3.5-27B INT4 as actor and Qwen3.5-4B as drafter, it matches sequential accuracy while cutting latency 10.23% over the Speculative Actions baseline and 18.59% over sequential execution on the Tau^2-Bench Telecom subset, and cuts AppWorld wall time 7.7% over SA and 44.9% over sequential with a small completion drop. Code is public.