HybridAgent uses a lightweight HybridLink to switch between latent-space reasoning and text-based tool calls, retaining multi-turn context through shared working memory. Across three model sizes and 11 benchmarks, the authors report a 10.4% performance improvement, 35.5–54.7% fewer tokens, and 1.8–2.6 times faster inference relative to comparison systems. The public model card shows that loading its Qwen3-based checkpoints with Transformers alone enables only text mode. These are self-evaluations of a dedicated training and execution system, not universal improvements that transfer directly to arbitrary agents.