IBM's CUGA runs 24 agent demos on 120B open weights — no frontier model needed
ibm released cuga, an open-source agent harness that uses planning and reflection to run complex apps on smaller open models instead of expensive apis, releasing 24 working examples to prove the architecture.
ibm released cuga (configurable generalist agent), an open-source agent framework designed to eliminate the boilerplate required for agentic applications like state management and tool wiring. the system topped the appworld and webarena benchmarks by using a 'codeact' approach and reflection steps to correct errors without re-deriving context, allowing it to run on smaller open models like gpt-oss-120b instead of relying solely on frontier apis. ibm demonstrated this capability by releasing two-dozen single-file applications ranging from movie recommenders to cloud architecture advisors.