LLMs are General Asynchronous Agents
George Yakushev, Denis Mazur, Vladimir Bartenev, Vyacheslav Zhdanovskiy, Timofey Byzov, Vladimir Kaurkin, Vadim Pastushenko
Yandex · Together AI · HSE University · Yandex School of Data Analysis
Published September 28, 2026 · on the feed September 30, 2026 · ▲ 77 · arXiv 2609.35427
The abstract
Modern LLMs are increasingly capable as autonomous agents, but they follow sequential interaction cycles: read, think, reply or call tools, repeat. Many real-world use cases are not sequential: voice assistants, embodied agents, and monitoring systems receive new inputs while they think or perform another task. Modern LLMs address this with specialized architectures for voice interaction and video streams, VLAs for robot control, asynchronous tool calling for API usage, and others. In this work, we generalize from different asynchronous tasks to general asynchronous agents that can adapt to different types of concurrency. To achieve this, we develop an asynchronous LLM framework that lets users (or the agents themselves) define inference coroutines with overlapping memory states. We showcase that Qwen 3.x models are capable of asynchronous operation for streaming video understanding, videogames, and monitoring, without task-specific training.
The Brief
1:41A short podcast summary of the paper. Written and voiced by AI.
Transcript
LLMs are General Asynchronous Agents. By George Yakushev and six others, from Yandex, Together AI, and two other institutions. Published on the twenty-eighth of September. Its code repository is worked out: about two hundred source files, with a readme and a licence.
Modern LLMs are increasingly capable as autonomous agents, but they work in strict turns: read, think, reply or call a tool, then repeat. Plenty of real-world uses do not wait their turn. A voice assistant, an embodied agent or a monitoring system receives new input while it is still thinking, or while it is busy with something else. Today each of these gets its own fix: specialized architectures for voice and for video streams, VLAs for robot control, asynchronous tool calling for API usage, and others besides.
The authors generalize across those separate cases. Instead of one specialized answer per kind of concurrency, they aim for general asynchronous agents that can adapt to different kinds of concurrency.
The vehicle is an asynchronous LLM framework. In it, users, or the agents themselves, can define inference coroutines whose memory states overlap. The description stops there: several strands of inference, defined by whoever needs them, sharing parts of their memory as they run.
The demonstration uses Qwen 3.x models. They operate asynchronously on streaming video understanding, on videogames and on monitoring, and they do so without any task-specific training.
Behind the title is a claim about where the capability lives. If an off-the-shelf model can run concurrently given the right framework around it, then asynchrony is less a property to train into each specialist system and more a matter of how inference is organized.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
The world doesn't wait for an agent's turn. Start with why that hurts.
You run language model agents that work in strict turns: read, think, reply or call a tool, repeat. That suits coding, but not a voice assistant interrupted mid-thought.
The road ahead
- 1 · The problem ~7 min Language model agents work in strict turns, but many real uses send new input while the agent is still thinking or acting.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Split one model's inference into concurrent coroutines that share memory directly, so existing models act asynchronously without task-specific training.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Define coroutines, give each a cache block and a view over the others' blocks, then run them batched on existing Qwen models over streaming input.
- 6 · The experiment ~6 min You’ll guess: To operate asynchronously on streaming video, videogames and monitoring, what did Qwen 3.x models need?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.