Hybrid Reasoning Agents: Which Monitor Still Works
A third of the steps go silent, the bill drops 2.7x, and the monitoring test only asked whether math answers were right.
- At
14Bparameters, HybridAgent's estimated generation cost drops from$470to$176per million questions (2.7x cheaper), as steps fall from about2,200to900per question. - About
35%of HybridAgent's steps at14Bare hidden, with visible text shrinking43%(from497to284tokens) while forward passes fall only6%(497to467). - In the paper's monitoring test on
560math problems, a strong checker (GPT-5.6-Luna) scored0.989AUROCon HybridAgent's traces versus0.990on a text-only agent's, while a weaker checker (GPT-4o-mini) scored0.860versus0.818. - HybridAgent outperforms the matched text-only baseline
ARPOby4%to6%on math/knowledge tests and27%to38%on agent benchmarks likeGAIAandWebWalker, while scoring only16.0versusKimi-K3's31.4on Humanity's Last Exam. - Goodfire's activation probes caught
3.1%more reward hacks than a chain-of-thought monitor onKimi K3, but7.9%fewer onGLM 5.2, and its cyber-misuse probes hit93%recall at a5.5%false-positive rate for under$200per million exchanges.
Tool-using AI agents write out every step of their reasoning as text before they search, run code, or answer, and every one of those words costs generation time and money.
HybridAgent, an AI agent from Stanford and Together AI, cuts that text down by doing part of its thinking without writing anything down.
What makes it interesting is the bill. At 14 billion parameters, the paper estimates the generation cost falls from $470 to $176 per million questions, 2.7x cheaper, because the agent finishes in about 900 steps per question instead of 2,200.
Written reasoning, the chain of thought an agent produces before it acts, is also one of the main ways anyone checks what an agent is doing, and HybridAgent writes less of it.
In the authors' own example, it's asked how many seconds sound takes to travel 10 kilometers. The trace shows one web search, two Python runs, and the answer, 29 seconds, while 138 of its steps in between were never written as words.
Hidden steps like these make up about a third of the work across the paper's tests at 14 billion parameters. For each one, the model feeds a long vector of internal numbers back into itself instead of writing a word.
It still writes some of its reasoning in plain text, and it always switches back to text to call a tool, so tool calls stay readable.
Its authors report that the shorter trace can still be monitored, based on a test that asked a second model one question about 560 math problems: was the final answer right? The code and data are open source.
So if your agents use tools, which check do you trust before you ship? And has anyone tested that check on the problem you fear: a wrong answer, or an agent doing something it should not?
The shorter trace drew pushback within hours. A governance researcher at the Machine Intelligence Research Institute replied to the launch post that the work undermines one of the few ways there are to monitor AI.
The same week, AI interpretability lab Goodfire published an essay arguing that a model's own neurons are far easier to read than the field had assumed, which points to a different way of watching an agent altogether.
So in this piece, we are going to discuss:
- the three checks you can run on your agent, with one worked case
- what you save in cost and what you lose in sight
- the checks Goodfire already runs inside models
- the paper's next step, and the warning in Goodfire's essay that it runs into
How HybridAgent reasons without writing
HybridAgent reasons without writing by feeding its own internal state back into itself, instead of turning each step into a word.
A language model normally ends each step by picking a token, the next word piece, and reading it back in. In what the paper calls latent mode, HybridAgent skips the token and loops its last internal state, a long vector of numbers, straight back in as the next input (Looped transformers: three different machines covers related looping designs).
It enters latent mode by writing a <latent> marker. From there, a small adapter called HybridLink converts the internal state into the form the model reads as input, for 16 to 256 steps at a time.
When a tool has to fire, it drops back into text, so tool calls and their results stay plain text in the trace, as the lead author explained in the launch thread.
What the paper adds is a text loop for tool calls around that hidden reasoning, across many turns. At 14 billion parameters, Coconut, an earlier method that loops hidden states with no text loop for tools, scores 6.8 on the GAIA agent benchmark against HybridAgent's 52.4.
Don't miss what's next in AI
Join 300,000+ engineers and researchers who get the signal, not the noise. Create a free account to read the rest of this story.
- Full access to in-depth AI research breakdowns
- Be the first to know what's trending before it hits mainstream
- Daily curated papers, repos, and industry moves