Nota reported joint optimization of the harness and inference path in an environment using Cline and the same model. In tests on four RTX Pro 6000 GPUs, it claims the UltraFast setting reduced task time by up to 47% and tokens by 60%, while the multilingual SWE score fell from 65.7% to 62%. The Fast setting was reported to cut time by 32% and tokens by 42%. These are company results for specific hardware and benchmarks, not universal improvements. Agent operations need to compare success rates under matching conditions alongside cost and speed.
Back to Tech
Nota reports agent optimization without changing the model
Maximum speed gains need to be assessed alongside changes in task success.