IBM Research says agent harness alone moved AppWorld task completion from 30% to 81%

IBM Research published a comparison of three agent harnesses on AppWorld (90 tasks) using the same model, GLM-5.3-flash: CUGA passed 81.1%, Hermes 46.7%, Deep Agents 30.0%. The CUGA team ran it and used a configuration the others did not; it is one benchmark with a small sample. CUGA is Apache 2.0.

HEAT

OUTLETS

0

FIRST SEEN

Nov 3, 2026, 11:05 ET

LAST SIGNAL

3m ago