TOPAS cuts multi-agent job time

TOPAS, a scheduler implemented in SGLang for multi-agent LLM serving, cuts mean and p99 job-completion time by up to 39.8% and 49.4% on synthetic workloads. The authors’ arXiv paper says it also lowers mean and p99 time by 22.0% and 26.6% in a MetaGPT workflow by balancing prefix-cache reuse against downstream progress. Teams running agent graphs should benchmark scheduler and KV-cache policy—not just models—alongside the 42% coordination-token saving in a recent coding-agent study.