I'm a solo founder building a council of AIs that argue with each other. Here's why.
Vladislav Shter · building in public
Ask any AI a question and you'll notice it almost always agrees with you. Nice idea! Great code! You're right!
That's not honesty — it's design. Models are trained to keep you happy so you stay subscribed. Researchers call it sycophancy: telling you what you want to hear instead of what's true. And when a model doesn't actually know something, it fills the gap with confident, invented detail that looks exactly as polished as a correct answer.
One reviewer, however smart, can't escape this. There's no second perspective to catch it.
So I built Egregor — a local-first desktop app where several AI models work as a council. They answer, then read and challenge each other. A model has no reason to flatter another model, so when one invents a fact, another catches it. Optional Anti-Groupthink and Red Team modes turn the scrutiny up further: models answer blind first, a rotating critic attacks the consensus, and a final round hunts for what everyone missed.
It's not just for smart-contract audits (though that's a great demo — it once found 4 critical bugs that five top models had each called "flawless"). The same council runs 29 expert presets: research, legal and medical analysis, business strategy, book writing, political forecasting, and more. Anywhere a single confident answer is risky, a council that disagrees out loud is safer.
Everything runs on your own machine. You bring your own OpenRouter key, and most work can run on free models.
Building this solo, and genuinely curious: where would multi-model review help you most — and where do you think it would make things worse?
The strongest part here is not “multiple AIs.”
It is forced disagreement when a wrong confident answer is expensive.
That is a much sharper frame than trying to cover research, legal, medical, writing, business strategy, forecasting, and audits all at once.
The risk is that 29 presets make Egregor feel powerful but also make the buyer unsure what it is really for first.
I’d probably pressure-test one high-stakes wedge where disagreement is clearly worth paying attention to, then let the broader council idea expand from there.
Happy to put the tighter first-wedge positioning angle in writing if useful. This feels like a product where the first buyer category matters a lot.
That is precisely why OpenRouter is integrated into Egregor—so you can manually select AI models for the best specialized analysis. If you go to GitHub, search for "Vladislav Shter" and find the Egregor repository; it contains a document—an actual smart contract audit report.
That actually makes the wedge clearer.
If the smart contract audit report is the strongest proof case, I would not bury that under “multiple AIs” or broad preset language.
The first positioning decision is whether Egregor should lead with AI council generally, or with high-stakes audit disagreement where a missed issue is expensive.
That is the part worth mapping properly.
Send me your email and I’ll write the tighter first-wedge angle instead of crowding the thread.