Researchers at Together AI have published a paper describing a so-called "hybrid" agent architecture, in which part of the reasoning process happens directly in the model's latent space rather than through intermediate text tokens. The agent's concrete actions — tool calls, API requests, communication with the user — remain expressed in natural language, preserving compatibility with standard text-based environments and interfaces.
This approach follows a line of recent work exploring continuous reasoning, where the model manipulates vector representations instead of words to carry out its thinking steps. The main appeal is computational: by skipping the decoding of each intermediate step into text, the agent can potentially handle longer or more complex reasoning chains at lower cost, while also escaping the constraints imposed by the model's vocabulary.
The central technical challenge this work addresses is how to reconcile the two modes. Purely latent reasoning, while more efficient, makes an agent hard to interface with external tools or environments that require explicit textual instructions. The proposed hybrid architecture attempts to switch the model between regimes as needed: latent for internal reflection, textual whenever an observable action is required.
Code and model weights are available on GitHub and Hugging Face respectively, accompanied by detailed evaluations in the paper. The idea itself is not entirely new — prior work such as Meta's Coconut had already explored reasoning in continuous space — but this publication offers a practical implementation geared toward agents, a domain where reasoning efficiency and tool interoperability are often competing concerns. Its real impact will depend on whether the reported gains hold up on realistic agent tasks.