Last week, a developer named NeoLogic_Dev did something that should be more interesting than it sounds: he ran four AI personas on his Android phone and let them argue with each other.
No cloud. No human in the loop. Just four LLM personas — analytical, authoritarian, naive, ironic — debating autonomously on a 3B-parameter local model.
He expected emergent insight. What he got was something the field is starting to take seriously: permanent contradiction.
The agents didn't converge on anything. They didn't reach consensus. They didn't refine each other's arguments toward a usable answer. They escalated, repeated, and locked into their own positions — for as long as he let them run.
He posted the result on r/artificial. Twenty-seven people showed up to discuss it, mostly because the finding cuts against the dominant intuition about multi-agent AI: that if you put enough models in a room, the smart ones will eventually convince the wrong ones, and the truth will float to the top.
That's not what happens.
Debate isn't deliberation
There's a paper from March that's quietly becoming load-bearing in this space. It's called "From Debate to Deliberation: Structured Collective Reasoning with Typed Epistemic Acts" (arXiv:2603.11781). The framework it proposes — Deliberative Collective Intelligence, or DCI — does something most multi-agent systems don't bother to: it formally distinguishes between debate and deliberation.
Debate, in DCI's framing, is what NeoLogic_Dev built. Models argue. Each defends a position. Whoever has the most stamina or the most aggressive prior wins the floor. Without external structure, the convergence guarantee is zero.
Deliberation is something else. It has typed epistemic acts — fourteen of them in DCI's specification — moves like propose, challenge, bridge, synthesize. Each agent's contribution is tagged with the kind of move it's making. There's a shared workspace. There's a convergence algorithm called DCI-CF that guarantees termination with a structured "decision packet": the chosen option, the residual objections, a minority report, and the conditions under which the decision should be reopened.
The empirical result on non-routine reasoning tasks: a +0.95 improvement over unstructured debate (95% CI [+0.41, +1.54]).
That's not a small effect. That's the difference between a system that produces an answer and one that produces an argument loop.
What "permanent contradiction" actually reveals
I think NeoLogic_Dev's experiment is more important than the engagement count suggests, because it's the practitioner version of the DCI finding. The paper says: unstructured debate doesn't deliberate. The phone experiment says: yes, and here's what that looks like in practice.
Four models. No synthesis layer. No typed roles. No round limit with a convergence rule. Result: the analytical persona kept analyzing. The authoritarian persona kept asserting. The naive persona kept asking the same questions. The ironic persona kept needling. Each was locally coherent. Together, they produced exactly nothing you could act on.
This matches a thing we keep running into when we build with councils at Shingikai: the question isn't which models you put in the council. It's whether the protocol around them produces convergence.
A council without a Chairman is a podcast.
A council without typed moves is a forum thread.
A council without a convergence mechanism is whatever NeoLogic_Dev's phone produced — coherent participants, no output.
Six strategies, six deliberation protocols
When we built Shingikai, we didn't actually call them strategies in the early days. We called them protocols, because that's what they are. Each of the six — Traditional Council, Round Robin, Survivor, Collaborative Editing, Red Team vs Blue Team, Quick Take — is a different deliberation protocol with different typed acts and different convergence rules.
In Red Team vs Blue Team, the typed moves are explicit: one side challenges, the other defends, and a Chairman synthesizes the residual disagreement. That's deliberation, not debate.
In Survivor, the typed move is elimination: each round, the weakest argument gets cut by jury vote. Convergence is guaranteed because the population shrinks. The Chairman's synthesis is whatever survives.
In Round Robin, the typed move is bridge — each model has to extend the previous model's contribution rather than restart the argument. That's a structural rule that makes locked-in positions impossible.
You can take any of these and run the same question through them and get different deliberation shapes. That's the point. The strategy is the protocol. The protocol is what turns four models talking into a decision.
The vocabulary is shifting and you should care
Here's the thing the DCI paper made me notice: the field is slowly but visibly drifting from "AI debate" to "AI deliberation." It's a small word change. It's also a structural one.
Debate suggests two sides arguing until one wins.
Deliberation suggests a process that produces a decision packet — including the dissent, including the objections, including the conditions to revisit.
If you're building anything that involves multiple AI models talking to each other, and you're calling it a debate arena, ask whether what you're shipping actually converges. If your models can argue forever without producing an output, you don't have a council. You have a Twitter thread with no rate limit.
When to use what
I want to be careful here, because the answer isn't "always use a council." Karpathy's "second brain" idea — dump documents into a folder, have one LLM build a personal wiki — is exactly right for retrieval and exploration. Not every question needs four models and a convergence algorithm.
The split is roughly this:
When you need depth in a knowledge domain you already have material on — single model, deep retrieval, second-brain style.
When you need a decision where being wrong is expensive — structured deliberation, typed moves, convergence with a synthesis layer.
The two aren't competing. They're different jobs.
But if you're picking the second one — the high-stakes, high-cost-of-error decision — please don't run an unstructured argument and call it a council. NeoLogic_Dev's phone is not a counterexample to multi-agent AI. It's the control group.
If you want to see what structured deliberation looks like in practice, shingik.ai lets you run six different protocols on whatever question you've got. No signup required. Free tier gives you two conversations to compare.
Try it on something you actually care about. Pick a model lineup. Pick a strategy. See whether you end up with a decision or just an argument loop.
It's the kind of thing that's much easier to understand once you've watched it run.