Hybrid agents cut reasoning cost while keeping tool use inspectable
HybridAgent lets an agent switch between latent reasoning, which happens internally, and text reasoning, which supports tool use and inspection. Across three model scales and 11 benchmarks, the paper reports 10.4% better performance, 35.5% to 54.7% fewer tokens, and 1.8x to 2.6x faster inference, with claimed cost savings of 2.7x to 3.7x. A HybridLink handles the transitions while continuous working memory preserves context across turns.