The team reports lower token use and faster inference across 11 benchmarks, while its cost savings remain estimates that exclude prompt processing and tool calls.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [arXiv](https://arxiv.org/abs/2610.18861)
Why it matters #
HybridAgent tests whether agents can keep tool access while moving some reasoning out of token-by-token text generation. Its cost estimates exclude prompt processing and external tools, so they do not establish production economics.
Together AI researchers led by Jiaru Zou released HybridAgent, a system that lets language-model agents reason in continuous latent space and switch back to text when they need to use tools or deliver an answer. In a paper submitted to arXiv on October 13th, the authors report fewer text tokens and faster inference than their tested baselines across 11 benchmarks. The results are the authors' own, not an independent production evaluation.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.