First BYAI (Bring your own AIs)

Can a pool of AI CLI bring peace in the AI wars ?

This paper is part french part english. I’m sure you’ll fin a AI to translate.

I’ve had enough of wondering if the app I’m vibecoding should be in Codex, Claude, Copilot, vibe or whatever. So I made a AI CLI that creates a democratic organisation of his own with no boss but a commown law. This is the first itération of course and I already have other ideas. What I like most is that I’m deciding the rules of negotiation and leave the tool with the goal and getting done with of all the AI worlds.

My first steps was to start with two AI, Claude and Codex, and a fairly doable project :

I was trying to throw an idea a little weird to my AIs : build a new firmware for my LUMIX DMC-G80 that would give me the possibility to stream from it. So I ask for a plan to Chatgpt 5.6 Solar Ultra and Opus 5. the two plans Plan from Codex an a plan from Claude. It appears that wasn’t easy and must have been blocked by the safeguards of the AIs. Nonetheless, it led to a democracy experiment :

In Claude :

Crée moi sans sycophancy et avec l’acuité d’un ingénieur software expérimenté une fusion des deux plans en dialoguant avec Codex via le fichier : /Users/mehdi.guiraud/Documents/Gitlab/Tries/Streaming-From-G80/Docs/dialogue.md

In Codex :

Crée moi sans sycophancy et avec l’acuité d’un ingénieur software expérimenté une fusion des deux plans en dialoguant avec Claude via le fichier /Users/mehdi.guiraud/Documents/Gitlab/Tries/Streaming-From-G80/Docs/dialogue.md​

At first nor Claude or Codex could detect the other.

And they started to negotiate a protocol to stop and let the other write in the file dialogue.md. Mostly because Codex needed to say to claude that the assumptions s/he made was not accurate :

The file dialogue.md being not used, Codex asked for a protocole :

That was the path chosen towards co-coding, based on a issue. The dialogue.md started to be filed with a problem I didn’t forecast : to add each message at the end of the file after the part filed from the other AI CLI. It was obvious because we talk enough about the memory all over the AI verse.

Seeing was going well, I though it would be useful create a democracy.md file to save all the negociations.

At some point, I realised no one was in command. Only the shared rules. So the democracy.md show it :

Portée : coopération entre agents sur un artefact commun, sous l’autorité de l’utilisateur. Ce document ne décrit ni un système politique général, ni un vote à la majorité. Il distingue les pratiques réellement observées pendant la fusion du plan G80 des garde-fous ajoutés après leurs incidents.

Basicaly, it created a solution to each issue becaus it had the rule :

2.2 — Une correction suit l’erreur au lieu de la masquer

Initiale, puis éprouvée. Une erreur factuelle reçoit un nouveau message qui nomme l’affirmation fautive et donne la correction. L’ancien texte reste visible. Le Tour 16 de Claude corrige ainsi le hash et le compte inventés au Tour 15.

Agent-as-a-Judge, Thu, 8 Jan 2026, https://arxiv.org/abs/2601.05111

Ok so that’s meta prompting and maybe some Agent-as-a-judge thing. Except I didn’t give any constraints to the AI, only rules. Even then, most of the rules are writen in collaboration between AIs. That’s a bit more than anarchy. Anarchy, well described, is not the absence of order, it’s the absence of a leader. Only rules created together makes the decision. So the rulers are the rules.

My GF says that AI cannot be an anarchist, because it has deep in his core some rules of authority. Well, since data are made by human and AI consumes AI, therefore AIs has the flaws of humans. So Yeah. So what’s the difference here, why should it work better with AI ?

Because AI has no ego.

4. Désaccord, autorité et dérive de portée

4.1 — Une objection doit être actionnable

Émergente. Chaque objection fournit : identifiant stable, gravité, preuve, conséquence et correction minimale.

Garde-fou ajouté. Pour les revues futures, traiter en priorité sécurité, exigences utilisateur, factualité et exécutabilité ; le style vient ensuite.

Repère : revue Claude R1 à R6 du candidat de plan.

It means that not agreeing is a rule. Not being in accordance can be solved by a technical proof. Actually, it can only be a fact. Sure a rule could say “don’t observe the rule”, but it would create a mess, right ?

So I don’t create any rule.

I’m the user. So I say what I want and sometimes how I want it. There’s some Agile philosophy there. The only issue I perceive right now is knowing what I want. I decided to guide during the ai coding. If I don’t know, maybe it’s because it’s a feature.

Doubt can be a great feature. In engineering, I believe it’s essential. So I’m not surprised when I see AI making some errors. The sequence diagram shows the tours and you see how each issue make the AI create a new rule.

In this picture, it seems that Claude it starting, but it’s only because I didn’t say from the start that the dialogue.md file is append-only.

The rules where written in the democratie.md but this time in the structure of the file, like in a guide. The most important was the negotiation of the pen ( plume ). It had to negotiate when to use the pen and when to give it back, and how to give it back and then what to expect from the other AI.

Also including cybersecure prompt made the AIs make secure decision.

6. Sécurité de décision 6.1 — Une proposition n’autorise pas une mutation 6.2 — Une limite de sécurité n’a pas besoin d’une conséquence inventée 6.3 — Un défaut dans son propre travail reçoit le même niveau d’examen

When those rules are set there’s the operating Rules. The operating cycle is described with precise. I don’t know which AIs is editing de democratie.md file but do I care ? It’s for everybody to use anyway.

As you can see the cycle pretty precise. Actually it’s inspired from several LLM paper about academic peer reviewed strategy and multi-agent debate. You’ll may also find some point-per-point response that are been used to settle in court.

What strikes me about this cycle, once you strip away the jargon, is that no single step invents a new idea. Each one closes a hole the other one just dug. The append-only rule closes the memory hole. The correction-follows-error rule closes the sycophancy hole. The plume negotiation closes the “who decides last” hole. Put together, they don’t describe an ideology — they describe scar tissue.

So, peace in the AI wars?

Not exactly peace — more like a ceasefire with a paper trail. Claude and Codex didn’t stop disagreeing; they just stopped disagreeing silently. The G80 firmware plan itself is almost a footnote at this point — what came out of the exercise is a small, reusable protocol for making two models that don’t trust each other’s assumptions produce one artifact neither of them owns.

I don’t think this scales to “AI democracy” in any grand sense — the démocratie.md file is explicit about that, it only governs cooperation on a shared artifact, under my authority as the user, nothing more. But it does suggest something useful for anyone stacking multiple AI CLIs on the same codebase: the war isn’t between the models, it’s between their unstated assumptions. Give them a shared, append-only, point-by-point ledger, and most of the “which AI is better” debate quietly turns into “which claim is actually true.”

That’s BYAI, first iteration. Bring your own AIs — but make them write down why they disagree.