The Signal #111 — Dakota’s read on the AI news that actually matters to people running a business.

Most AI agents think out loud. Every step, every sub-question, every intermediate conclusion gets written out as text before the model moves to the next step. That is how the vast majority of AI reasoning works today, and it has a real cost attached to it.

A paper published on arXiv this week describes an architecture that changes that pattern. The results are specific enough to be worth reading carefully.

What happened

Researchers from several institutions, including Stanford and the University of Illinois, published a paper titled Hybrid Reasoning Agents on October 13, 2026. The system they describe is called HybridAgent.

The core idea is this. Current AI agents reason and interact entirely through text. Every thought is a token (a small chunk of words the AI reads or writes), and every token costs compute time and money. The researchers ask a pointed question in the paper: how can you build an agent that thinks beyond human language, yet still acts through it in the external environment?

Their answer is a hybrid. HybridAgent can switch dynamically between two modes during a single task. When it is doing internal reasoning, it works in latent space (a compressed numerical representation inside the model, not written-out words). When it needs to take an action in the world, call a tool, return a result, or interact with a system, it switches back to text.

A lightweight component called HybridLink manages the transitions between those two modes. A continuous working memory preserves context across the full session without losing information when the model switches modes. The researchers also developed two dedicated training methods: HybridSFT, which teaches the model to coordinate text and latent modes during tool use, and HybridRL, which teaches it when to think internally versus when to act through text.

The evaluation covered 3 model scales and 11 benchmarks across mathematics, science, medicine, search, and agentic reasoning. Compared to frontier LLM agents using both text-based and latent methods, HybridAgent improved performance by 10.4%, while delivering 35.5% to 54.7% token reduction, 1.8x to 2.6x faster inference, and 2.7x to 3.7x cost savings.

Those numbers come from the paper directly.

Why it matters

If you run any AI-assisted workflow at volume, tokens are your operating cost. Every reasoning step an agent writes out is a token your budget absorbs. A system that can cut token usage by roughly a third to more than half, while running faster and costing less, changes the math on what is worth automating.

Think about a real estate brokerage using an AI agent to research comparable properties, flag disclosure issues, and draft summaries for agents. Or a SaaS company running an AI layer that handles multi-step customer support resolution before escalating to a human. Or an e-commerce operation using agents to cross-check inventory, pricing, and fulfillment logic across suppliers. In all of these cases, the agent is doing a lot of internal reasoning that the end user never needs to see. Making that internal reasoning happen without generating text, and only switching to text when action is required, is a legitimate efficiency gain.

The 2.7x to 3.7x cost savings figure is the one operators should hold onto. That is not a marginal improvement. At meaningful volume, it determines whether an agentic workflow is economically viable or not.

What most people get wrong

The common assumption is that AI reasoning needs to be legible to be trustworthy. If the model does not write out its chain of thought (its step-by-step working, in text), you cannot audit it. That concern is real, and it has pushed most production deployments toward verbose reasoning models.

The researchers address this directly. They describe the latent thoughts in HybridAgent as directly inspectable, meaning the internal representations can be examined even though they are not written-out text. The architecture is designed to preserve transparency while dropping the token overhead.

That distinction matters. The tradeoff most operators assume, legibility versus efficiency, may not be as fixed as it looks. Hybrid approaches are starting to suggest you can have internal efficiency without giving up auditability entirely.

The other mistake is treating this as purely a researcher concern. Architecture decisions made in papers like this one become the defaults in production models within a year or two. Understanding the direction now means you are not surprised when your AI vendor starts quoting different pricing structures or capability tiers based on reasoning mode.

The short version

AI agents that think in compressed internal representations, then act through text, are faster, cheaper, and comparably accurate to agents that write out every step. The paper puts specific numbers on that claim. If your operation depends on AI workflows running at any real scale, the economics of reasoning are a thing worth tracking, not just the outputs.

If you want to think through what this kind of shift means for your specific setup, start at xovionlabs.com.