Speculative Macro Commit for Faster Tool-Using Agents

Zeyu Liu, Souvik Kundu and Peter A. Beerel · accepted at MLSP 2026; arXiv 3 Sep

The authors pair an authoritative actor with a faster speculative drafter operating on an isolated snapshot. They report latency reductions of 18.59% versus sequential execution on a telecom subset and 44.9% on AppWorld, with a small task-completion reduction on AppWorld. Practical value: lower tool-agent latency. Limits: author-reported results on selected benchmarks; no independent reproduction. Paper