The smallest one in the room is in charge

Oct 12, 2026

← Blog

You don't clear a raid with five tanks. You bring a tank, a healer, somebody who hits hard, and one person who yells "not yet!" before every pull.

AI models are the same. One writes tidy code. Another is good at math. A third finds the bugs the first two left behind. So why does everybody keep sending a single hero to solo the whole dungeon?

Two papers from Sakana AI, both from ICLR 2026, are about exactly that: how to make a party out of models. They are called Trinity and Conductor, and I had a lot of fun reading them. Here is how they work. No math, I promise.

Trinity plays turn-based

Picture a tactics game. Three units, one task, five turns.

Every turn somebody has to answer two questions: who moves, and what do they do. In Trinity a unit can do only three things:

  • Think. Step back and come up with a strategy.
  • Work. Actually do the next step.
  • Verify. Look at the result and say "done" or "nope".

If the verifier says "done", the game is over. If it says "nope", the next turn starts. After five turns, time is up.

Now the funny part: who is the player? Not a big model. A small one reads the board, and the piece of it that actually learned how to lead is about ten thousand numbers. That fits on a floppy disk, with room for thirty more copies.

And it was trained like a roguelike. A run ends with a win or a death, nothing in between, and that is all the feedback there is. Ordinary training chokes on that: too little signal. So they bred it instead. Thousands of runs, the best ones survive, get mixed, and off we go again.

In the end it beat every model in its own party on coding tasks. A raid leader the size of a floppy disk. I love it.

Conductor writes the strat before the pull

Conductor is a different kind of leader. It doesn't decide turn by turn. It looks at the task once and writes the whole plan, in plain words, the way a raid leader types the strategy into chat.

Three lines for every step:

  • who does it,
  • what exactly they should do,
  • what they get to see of everybody else's work.

That third line is my favourite. It is fog of war, on purpose. The reviewer doesn't need the whole brainstorm, and the one writing code doesn't need three opinions about style. Everybody sees just their piece of the map.

Nobody taught Conductor any tactics. It got a reward only when the final answer was right. And like players in any game, it found the meta by itself:

  • a rat in the cellar gets one model and one shot;
  • a boss gets the full party: one plans, one builds, one reviews;
  • sometimes it asks a few models separately and then lets them argue.

Five steps at most, and on average three were enough. No grinding for the sake of grinding.

It can even write itself into the plan. Then it watches the replay and, if the party wiped, comes up with a new strat.

So, turn-based or strat?

Trinity reacts, Conductor plans. Different genres, same lesson: the leader doesn't have to be the strongest unit. It has to know its party. And it knows the party from results, not from a config somebody typed by hand.

What I'm taking home

I'm building Levocat Flock on the same idea, so here is my loot:

  1. Don't solo. Different models, different classes.
  2. The one who builds doesn't review. A second pair of eyes should be a different pair.
  3. Fog of war is a feature. Show each model only what it needs.
  4. Don't gather a raid for a rat. Easy things go to one model.
  5. The lead is earned. It is measured on real tasks, not appointed. Flock stands on this one.

The papers, in case you want the math I promised to skip: Trinity and Conductor.