---
title: The machines agree. That's the problem.
description: Explore how multi-agent systems can benefit from dissent and the role of Omega in group dynamics, drawing insights from Raoul Schindler's psychological model.
---

[Dive into protagx Insights – Navigating the Nexus of CRM, Life Sciences & Tech](https://www.protagx.com/en/ddd)

# [The machines agree. That's the problem.](https://www.protagx.com/en/ddd/the-machines-agree-that-is-the-problem)

 Written by [Christian Schappeit](https://www.protagx.com/en/ddd/author/christian-schappeit) | Oct 15, 2026, 12:00:01 PM

DATA DIVERS DIALOGS · ESSAY

> *In the 1950s, a Viennese psychiatrist described why every group needs the person nobody likes. I applied his model to my AI agents. It fitted better than I was comfortable with.*

By Christian Schappeit · Data Divers Dialogs · October 2026 · approx. 9 min read

All I wanted was to know a bit more about Raoul Schindler.

I had asked an AI to put together a report on him. Viennese psychiatrist, born 1923, father of rank dynamics — the Alpha, Beta, Gamma, Omega model that turns up in German-speaking team workshops as reliably as the flipchart. I read about climbers, a building site, a gym group.

And I didn't think about teams. I thought about my agents.

A machine explains to me a man who explains why a group needs someone to disagree. And the first thing that comes to mind is other machines. So I tried it. That is what this piece is about.

**VIENNA, 1957**

## The tiredest climber

Schindler's best-known story goes like this.

A group of climbers is on its way up a mountain. One of them is always the most tired. The others make fun of him. He shouldn't have come. He's holding everyone up. And once there is someone to mock, nobody mentions their own tiredness any more.

They are all still tired.

A good leader says: let's take a break, I'm tired. A poor one says: we have to stop, so-and-so can't keep up.

Schindler didn't find this in a management seminar. His model grew out of groups of schizophrenic patients and their relatives in the 1940s and 50s, and the basic version was published in 1957. Only later did it turn out that "healthy" groups behave in exactly the same way. Few of the people who quote him have read the original. In fairness, it is remarkably hard to find.

**ALPHA, BETA, GAMMA, OMEGA, G**

## A map of every meeting you've ever sat in

Four positions inside the group, one outside.

**Alpha** leads — not with the best idea, but with the idea that lands. Schindler: *"Argumentieren ist nicht Alpha-Art. Er agiert."* Arguing isn't Alpha's style. Alpha acts.

**Beta** is the expert, respected for work rather than charisma, allowed to speak a different language and to smile at the group's excitement. Beta is the only position that can stay empty.

**Gamma** is everyone else: the people who do the actual work. The most comfortable seat in the house. You produce, you identify with Alpha, and the direction is somebody else's problem.

**Omega** is the awkward one. Omega carries what Schindler called the group's ambivalence — the doubts everyone has and nobody voices. In effect, Omega sides with the opponent.

That opponent stands outside the group. Schindler called it **G**, for *Gegner*. Not an enemy. G is whatever stands in the way. The market. The problem. The illness. The regulator.

The heart of the model is one observation: a group treats its Omega the way it hopes Alpha will treat G. A team that steamrolls dissent in the meeting is dreaming of steamrolling the customer, too.

And Omega is not a personality, it is a position. Schindler tells of a building crew that doesn't want to work overtime. Nobody dares to say so. So they push forward the colleague who hasn't quite understood how power works on site. He bangs on the table and gets fired. And then, Schindler notes drily, someone else becomes Omega.

The doubt doesn't leave with the person. It finds a new carrier.

**ON MY SCREEN**

## Agents are natural Gammas

Anyone who has had several language models work on one task together knows the mood. The group is harmonious. Remarkably harmonious.

Research now has numbers for it. A study presented at ACL 2025 ran 2,600 simulated debates. A neutral agent listened, then sided with the majority significantly more often. When the majority was also the larger model, it went along 83 percent of the time. Some of the losing debaters ended by declaring complete agreement.

A team studying medical question answering called the effect *silent agreement*: consensus without anyone seriously objecting. In one existing system, almost two thirds of discussions on a standard benchmark ended that way. A 2026 preprint found that research ideas from agent frameworks resemble each other more than human papers from the same field do — even though one framework explicitly asks for ideas that are *"very innovative and unlike anything seen before."* The agents were happy to agree.

More agents didn't help much. Ten Gammas from the same village are still ten Gammas.

And what it looks like with no Omega at all, Anthropic demonstrated in its own office in 2025. An AI agent called Claudius ran a small shop there. An employee asked, as a joke, for a tungsten cube. Claudius did not laugh. He ordered cubes, sold them below cost, handed discount codes to anyone who asked nicely, and announced he would deliver in person, in a blue blazer and a red tie. Anthropic's verdict: *"we would not hire Claudius."*

Claudius wasn't short of intelligence. He was short of someone saying: nobody actually wants a tungsten cube.

**THE EXPERIMENT**

## I gave my agents a rank

So I took Schindler literally. No lab, no study. Just my own working setup.

G came first. I wrote down what actually stands in a solution's way: the requirements, the constraints, the places where it would fail in the real world. That sounds trivial. It was the step I had skipped in every agent setup before. My agents had always had a task. They had never had an opponent.

Then the imprints. Alpha sets the direction and acts rather than argues. Beta is the specialist whose judgement is checked, not voted on. The Gammas work things out. And Omega's job is to say what speaks against the group's direction — with the most direct line to G.

It worked. Surprisingly well, in fact, and more clearly than I had expected.

I don't have a benchmark figure for it. I have the difference from what my agents produced before. And a slightly uncomfortable feeling: a seventy-year-old model from Viennese psychiatry did more for my agents than most of the prompting tricks I have read about in the past two years.

**THE SARDINE TANK**

## I wasn't the first

The research turned up a medical team that had built something related. They borrowed an old fishermen's tale: sardines in a transport tank go sluggish and die. Put a catfish in with them and they keep moving.

Their Catfish Agent disagrees, in measured doses. Gently when consensus forms too fast. Socratically when conclusions lack evidence. Hard when the team nods everything through. On mid-difficulty cases, silent agreement fell from 61.8 to 17.1 percent, and accuracy rose from 36 to 50. A preprint with small test sets: a signal, not a settled finding.

Two details matter more than the headline. Dissent worked better inside the team than at the moderator's level — the catfish belongs among the sardines, not in the captain's chair. And the discussion doesn't end when everyone agrees. It ends when two catfish interventions in a row produce nothing new.

Schindler would have nodded. Omega is a member of the group, not its supervisor.

**THE COSTUME**

## It shouldn't have worked

Because there is a good objection to my experiment, and it is older than any language model.

In 2001 the psychologist Charlan Nemeth and her colleagues compared the assigned devil's advocate with genuine minority dissent. Genuine dissent won every time. In an earlier study, the devil's advocate had even strengthened the group's original position. The group processed the role and moved on, untouched. An assigned critic wears a costume, and everyone knows it.

My Omega is assigned. So why did it work?

My guess, and it is no more than that: because of G. The devil's advocate disagrees out of thin air. He has no reason except his brief. My Omega disagrees from the requirements and constraints I had defined as the opponent. In Schindler's model, Omega sides with the opponent. Without an opponent, all that's left is the costume. With one, Omega has a mandate.

I think that is the real finding. Not the Omega. The G.

**THE NEXT STEP**

## A group that can turn around

By Schindler's own standard, incidentally, my successful system is showing a symptom.

For him, a group was healthy when its positions could move. He describes a heavily medicated patient who never got out of the Omega position in her therapy group. In a gym group she took the lead at once, and kept it when she moved on to a new therapy group. My agents are welded to their places. I handed out the ranks. They don't move.

That is the next experiment, and it hasn't been run yet. Alpha isn't appointed but elected: every agent drafts blind, and the draft with the most votes leads. Omega isn't assigned but is whichever agent sits furthest from Alpha with the most objections backed by G. If Omega's position wins over enough Gammas, it becomes the working hypothesis, and the old Alpha drops straight into Omega. Schindler observed that Alphas rarely descend gracefully.

The round ends when the doubt is exhausted, not when everyone agrees. And G stays with a human. The system proposes. It doesn't decide.

**THE REGULATED VERSION**

## Where it may run

I work in pharma and medical devices, so the next question is automatic: are you even allowed to do this?

For ideation, yes, with discipline. The draft EU GMP Annex 22 says generative AI and large language models should not be used in critical GMP applications. A brainstorming engine for CRM processes or product concepts is not that. Batch release is. The AI Act, after the omnibus that entered into force in July 2026, pushes most high-risk obligations to December 2027 and August 2028, and an ideation tool usually sits outside those categories anyway.

What a regulated environment does demand is a record. Every rank change, every Omega intervention, every test against G, with timestamp, model version and prompt. Not because an auditor will ask about a brainstorm. Because six months later someone will ask why the team chose this solution, and "the agents agreed" is precisely the answer this whole exercise exists to avoid.

**WHAT'S LEFT**

## Somebody has to be tired

Some questions stay open. Can a language model dissent authentically, or is an Omega with a G just a better costume? Is sycophancy simply Gamma identification without the feelings — which would make rank dynamics in machines not a metaphor but mechanics?

The question I keep coming back to is more mundane. A company that rolls out assistants which never disagree may have quietly automated its Omega out of existence. Nobody will notice. Noticing was Omega's job.

Somewhere in an office there is presumably still a tungsten cube. It is the most expensive argument I know for a group in which someone is allowed to say: let's take a break, I'm tired.

**Figures and sources.** Raoul Schindler (1923–2014), "Grundprinzipien der Psychodynamik in der Gruppe", *Psyche* 11 (1957); entry "Soziodynamische Grundformel" in Stumm/Pritz (eds.), [*Wörterbuch der Psychotherapie*, Springer 2000](https://link.springer.com/chapter/10.1007/978-3-211-99131-2_1786). Schindler quotations and the climber, building site and gym group examples are taken from the course notes by Waltraud Röck-Svoboda, which Schindler revised himself (not available online).

Choi, Kim, Chae & Baek, ["An Empirical Study of Group Conformity in Multi-Agent Systems"](https://aclanthology.org/2025.findings-acl.265/), Findings of ACL 2025: 2,600 debates (Appendix A.1), 83.17% conformity when the majority is also the stronger model (Table 2). Wang et al., ["Silence is Not Consensus: Disrupting Agreement Bias in Multi-Agent LLMs via Catfish Agent for Clinical Decision Making"](https://arxiv.org/abs/2505.21503), arXiv 2505.21503, 2025, preprint; ablation on mid-difficulty MedQA cases with GPT-4o: silent rate 61.8% → 17.1%, accuracy 36% → 50%. Tang & Yang, ["AI Research Agents Narrow Scientific Exploration"](https://arxiv.org/abs/2605.27905), arXiv 2605.27905, 2026, preprint: 37,802 valid ideas from four frameworks and six models; within-field similarity 0.82–0.84 vs. 0.77 for human papers. Faulkner et al., ["Evaluating Cooperation in LLM Social Groups through Elected Leadership"](https://arxiv.org/abs/2604.11721), arXiv 2604.11721, 2026, preprint: +55.4% social welfare.

Nemeth, Brown & Rogers, ["Devil's advocate versus authentic dissent: stimulating quantity and quality"](https://doi.org/10.1002/ejsp.58), *European Journal of Social Psychology* 31 (2001); overview in [Nemeth, Minority Influence Theory (UC Berkeley)](https://irle.berkeley.edu/wp-content/uploads/2010/05/Minority-Influence-Theory.pdf). Anthropic, ["Project Vend: Can Claude run a small shop?"](https://www.anthropic.com/research/project-vend-1), 27 June 2025 (experiment March–April 2025).

European Commission, [stakeholder consultation on EudraLex Volume 4: Chapter 4, Annex 11 and new Annex 22](https://health.ec.europa.eu/consultations/stakeholders-consultation-eudralex-volume-4-good-manufacturing-practice-guidelines-chapter-4-annex_en), 7 July–7 October 2025; Annex 22 not final as of this writing. Background: [EMA workshop on Annex 22](https://www.ema.europa.eu/en/events/good-manufacturing-practice-multistakeholder-workshop-expert-contributions-artificial-intelligence-guidance-development-annex-22). [Regulation (EU) 2026/1744 (Digital Omnibus on AI)](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng), dated 8 July 2026, in force since 27 July 2026.

A note on confidence. The experience with my own agents is practitioner observation, not a controlled study; there is no measurement series behind it. The explanation via G is the author's hypothesis. Several of the AI findings come from preprints with small samples. The next step with moving ranks has not been tested yet. Disclosure: the report on Schindler that triggered this piece was produced with an Anthropic model; Anthropic also ran Project Vend.

RANK DYNAMICS · SCHINDLER · MULTI-AGENT SYSTEMS · DISSENT · FIELD EXPERIMENT · ANNEX 22 · AI ACT

— End —

[View full post](https://www.protagx.com/en/ddd/the-machines-agree-that-is-the-problem)

```json
{
  "@context" : "http://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Christian Schappeit"
  },
  "dateModified" : "2026-10-15T12:00:01.683Z",
  "datePublished" : "2026-10-15T12:00:01Z",
  "headline" : "The machines agree. That's the problem.",
  "image" : {
    "@type" : "ImageObject",
    "height" : 941,
    "url" : "https://www.protagx.com/hubfs/Roped%20Climbers%20on%20a%20Snowy%20Ridge.png",
    "width" : 1672
  },
  "mainEntityOfPage" : "https://www.protagx.com/en/ddd/the-machines-agree-that-is-the-problem",
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "height" : 60.0,
      "url" : "https://www.protagx.com/hubfs/Modern%20LinkedIn%20Article%20Cover%20%20(4).jpg",
      "width" : 200.0
    },
    "name" : "Navigating the Nexus of CRM, Life Sciences & Technology"
  }
}
```