# Hybrid Reasoning Agents Hybrid Reasoning Agents (HybridAgents) are [[AI Agents|LLM agents]] that think in [[Latent Reasoning|latent space]] and act through text. That's the whole idea in one line: the model reasons in hidden vectors (fast, cheap, unreadable), then switches back to normal tokens when it needs to call a tool or answer the user. The paper was published on arXiv on 13 October 2026 by Jiaru Zou, James Zou and colleagues, with the code under Together AI's GitHub organization. Why hybrid? Pure latent reasoning has two problems. Tools expect text (a JSON tool call isn't a vector), and nobody can read what the model is thinking. Text-only agents have the opposite problem: every thinking step costs tokens. HybridAgent tries to keep the best of both. ## How it works - **HybridLink:** a small component that handles the switch between latent mode and text mode inside a single generation - **Persistent working memory:** keeps the mixed latent and text context intact across turns, so the agent doesn't lose its thinking between tool calls - **Two training steps:** HybridSFT teaches the model to work with HybridLink, then HybridRL ([[Reinforcement Learning (RL)|reinforcement learning]]) teaches it *when* to think latently and when to act through text - **"Inspectable" latents:** each latent step can be mapped to its nearest token embedding, so you get a rough readable trace ## What the authors claim Tested at three model scales on 11 benchmarks (math, science, medicine, search and agentic reasoning): - About 10.4% better accuracy on average than the systems they compare against - 35.5% to 54.7% fewer tokens - 1.8x to 2.6x faster inference - 2.7x to 3.7x cheaper If that holds, it's a big deal for [[AI Agent Harness|agent harnesses]], where thinking tokens are a large part of the bill. ## What the Hacker News discussion added The thread was skeptical, and I think rightly so: - **What's new here?** The top comment pointed out that continuous chain of thought has been studied for about two years, and that tool calls already work by leaving thinking mode and emitting normal tokens. Their verdict: "I'm not seeing a win here" - **Monitorability is the elephant in the room.** Most of the discussion was about safety. A readable chain of thought is one of the few ways to oversee a model, and mapping vectors to their nearest token throws most of the information away. Worse, the paper suggests training against the latent reasoning, which is exactly the kind of pressure that could teach a model to hide its intent. One commenter noted the paper doesn't discuss that tradeoff at all, and that acknowledging it would have been more honest - **The released models didn't impress.** Early users ran the 8B and 14B models through [[llama.cpp]] and saw them loop and fail on simple questions and a classic logic puzzle. Part of that may be missing inference support, since the authors ship their own inference code. Still, "it only works with our code" isn't a great sign ## My take The numbers are attractive, but the benchmarks are the authors' own selection, and the first independent tests went badly. I'd wait for replication. The bigger question is the tradeoff. I run agents for hours on my own machine, and reading what they planned is part of how I trust them. Cutting thinking tokens by half isn't worth much to me if I lose that. The one thing I like in the design: actions stay in text. Whatever the model thinks, what it *does* (tool calls, file edits, answers) stays readable and checkable by the harness. If latent reasoning ever goes mainstream, that's the line I'd want to keep. ## References - Zou et al., "Hybrid Reasoning Agents" (arXiv, 13 October 2026): https://arxiv.org/abs/2610.18861 - Code: https://github.com/togethercomputer/HybridAgent - Hacker News discussion: https://news.ycombinator.com/item?id=50145337 ## Related - [[Latent Reasoning]] - [[AI Reasoning Models]] - [[Chain-of-Thought (CoT) prompting]] - [[AI Agents]] - [[AI Agent Harness]] - [[AI Tool Use]] - [[AI Safety]] - [[AI Post-Training]] - [[Reinforcement Learning (RL)]] - [[Embeddings]] - [[llama.cpp]] - [[Hacker News]]