# What if we could improve reasoning by having multiple AI agents discuss a problem in parallel? — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/3d4c03f813568d6d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-09 14:09 ingested 2026-11-10 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xrsleb/what_if_we_could_improve_reasoning_by_having/)

## retained text



```
Reasoning models spend a huge number of tokens on thinking to improve the accuracy of a single response, right?

The problem is that this happens sequentially in a single stream, so the longer the thinking process, the longer we have to wait for an answer.

So what if, instead of relying on a single long reasoning process, we sent multiple concurrent requests to the same model and had the responses interact with each other, kind of like a group of people discussing and solving a problem together?

We'd probably also need agents to consolidate the different answers, review them, and so on.

Luckily, we already have vLLM, which is great at handling multiple concurrent requests.

Has anyone experimented with this approach? If so, did you learn anything interesting?

What got me thinking about this was that I was getting 88 tokens/s with a single request, but 2,000 tokens/s in aggregate with 128 concurrent requests. It made me wonder: what if we could put all that extra throughput toward improving intelligence?
```



document fetched 2026-11-10 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xrsleb/what_if_we_could_improve_reasoning_by_having
