IBM reports that with GLM-5.3-flash fixed across 90 AppWorld tasks, CUGA passed 73, Hermes 42 and Deep Agents 27. The number of scenarios passing all three variants was 20, eight and two respectively. CUGA used decomposition, focused tools, pre-execution verification and reflection. However, harness configurations differed and this was a small sample on one benchmark. Policies were not enabled in the tests, so the results support workload-specific harness selection, not universal superiority or proven approval-rule safety.
Back to Tech
IBM reports 81.1% AppWorld task completion with CUGA using a fixed model
The 90-task comparison neither ranks every harness universally nor validates policy enforcement.