Same Model, Different Harness: CUGA, Hermes, and Deep Agents — When to Pick CUGA? | AI Archive

Same Model, Different Harness: CUGA, Hermes, and Deep Agents — When to Pick CUGA?

huggingface · 2026-11-03T15:57:39+00:00

IBM Research conducted a controlled experiment comparing three agent harnesses—CUGA, Hermes, and Deep Agents—using the same GLM-5.3-flash model on 90 AppWorld tasks. The study revealed that changing the harness significantly impacted performance, with task completion rates ranging from 30% for Deep Agents to 81% for CUGA. While Hermes and Deep Agents excel in personal assistance and long multi-step research respectively, CUGA is specifically designed for enterprise environments requiring strict rule enforcement. CUGA achieves higher consistency by integrating features like task decomposition, code execution sandboxes, and pre-execution verification. Crucially, it moves business rules out of system prompts into runtime-enforced policies, ensuring compliance regardless of model decisions. This approach allows for secure deployment within sovereign clouds or air-gapped networks while maintaining measurable cost controls. The article concludes that selecting the right harness depends on the specific workload, recommending CUGA for complex, rule-bound enterprise automation.