**Five diverse AI agents vs. five clones for 14 nights: a constant tied both**

> Five diverse AI agents vs five clones, 14 nights, 4,130 scored forecasts. The lenses held apart. A base rate I got 15x wrong tied both to a constant. I Ran Five Diverse AI Agents Against Five Clones for 14 Nights. A Number I Made Up Decided the Result. 4,130 scored forecasts, pre-registered. The lenses really thought differently. A constant tied both. Pawel Jozefiak Sep 11, 2026 5 1 Share The scoreboard I opened on the morning of 5 September said my thesis had won. Five AI agents with genuinely different lenses had beaten five identical clones on 9 of 14 nights, and their pooled Brier score was 0.0225 against the clones' 0.0275. Lower is better. I had wanted that table since the first night. I did not publish it. The reason is this post. Most of my last few months went into agents that work together: several of them, with different jobs, handing work to each other, with rules about what each one is allowed to see. At some point I stopped and asked the only question that matters there. When you add the second agent, what did you actually add? If it is the same model reading the same context, you added compute. More tokens, more latency, a second opinion drawn from the same well. If it knows different things and looks through a different lens, you added something else, and that something else is the whole case for building a network instead of buying more of one model. So I wrote the claim down before I could soften it. A network of agents compounds context, not compute. Agents with deliberately different roles, contexts and information diets should beat the same number of clones of the same model, at equal budget, on a task with real answers. If they tie, the "more agents means more intelligence" story is wrong, and I would rather learn that from my own data than from a bill. Fourteen nights later I have an answer. It is a fourth ending I never wrote down, and the thing that decided it was a sentence I wrote myself. Sold hard, measured rarely Look at how multi-agent gets sold right now. Panels, councils, swarms, debate rounds, a "team of specialists". The pitch is always that the group knows more than any member. Underneath most of those products sits one model wearing several system prompts, running at several times the cost of one call. I have been tracking who is actually winning in that market since February , and the panel pitch is in nearly every listing. I could not find anyone who had put the two arms side by side at equal budget, on a task with a scoreboard, and published the number including the boring one. It is an awkward experiment to run if you are selling the thing. So I built the scoreboard instead of arguing about it. The rules, written before anything ran These went into the plan first, and they are the only reason the rest of this post is worth reading: Two arms, one variable. Five clones against five diverse agents. Same model (Opus 5), same budget, same tools, same task, same morning. The only thing that changes is what each agent knows and how it is told to look. Reality grades it. No model scoring another model's work, because a language model judging language models is the same well again. Deterministic Python, public APIs, zero human judgment in the loop. Pre-registered. Scoring rule, pass threshold and the three possible endings written down before the first sample. A failed check is a finding, never a reason to edit the rule. The control is built strong. The clone pack is a full theory of why things spread, applied from first principles. A rigged control proves nothing and I would know I rigged it. Every experiment pays either way. A null result still has to leave me with a working tool. Nulls get published. Including the ending...

**来源信息**
- **来源**：Hacker News：AI 热帖
- **分类**：ai-models
- **发布时间**：2026-09-11 21:04（北京时间）
- **原文**：[打开原文](https://thoughts.jock.pl/p/five-lenses-one-brain-agent-diversity-experiment-2026)