StarHarness: Evolving Harnesses with Stratified Search for Enterprise Environments
StarHarness improves enterprise agent performance by evolving the harness around a fixed model.
The framework changes prompts, task framing, tool interfaces, skills, MCP providers, subagents, and loop settings without updating model weights. In tests on ITBench SRE, EnterpriseOps-Gym ITSM, and AutomationBench Finance, it beat the default harness by 20-35 percentage points after 4-12 accepted changes per environment. The paper says the gains held on tasks excluded from evolution and transferred across GPT and Qwen model families. HF Daily Papers' note
The framework changes prompts, task framing, tool interfaces, skills, MCP providers, subagents, and loop settings without updating model weights. In tests on ITBench SRE, EnterpriseOps-Gym ITSM, and AutomationBench Finance, it beat the default harness by 20-35 percentage points after 4-12 accepted changes per environment. The paper says the gains held on tasks excluded from evolution and transferred across GPT and Qwen model families. HF Daily Papers' note
score 5