EvoHarnessBench

"Can Your Agents Keep Pace with an Evolving Harness?" Salesforce Research, with UNC and UW–Madison, puts the non-stationarity in the harness rather than the task stream: 17 multi-stage harness streams over 802 tasks in which the available tools (520), skills (42), and agents (62) change under the agent. Authors Zixuan Ke, Vaidehi Patil, and colleagues; no dedicated benchmark repository yet beyond the MAS-Orchestra project.