I Trained a Decision Model to Play Super Mario Bros.

I trained a decision model to play Super Mario Bros. Linux laptop. Qwen 3.5, two billion parameters, a small decision head on top.

Mario faces what every agent faces. Partial state. Moving parts. Enemies, gaps, timing a single frame doesn't reveal. At each decision, he has to pick a move and whether to jump. An enterprise agent lives in the same loop. Incomplete context, shifting conditions, a bounded set of tools. Perceive, choose, act, observe the outcome.

The difference: Mario's choices are closed. A fixed set of moves, a score, a consequence. That is exactly what you want from any decision you hand a model. So the model scores the available moves. No generated text and no parsed explanation, just a choice.

Imitation learning gave the head a starting policy. Then reinforcement learning took over: reward for progress, penalty for dying. The backbone stays frozen during reinforcement learning.

The enterprise lesson is the boundary. A learned model can score how risky a payment looks. It must not decide whether the agent may make it. Policy defines how that score affects execution. Below the threshold, an otherwise authorized action proceeds. At or above it, additional review. The model produces the score. The organisation owns the threshold, versions it with the model and audits both. No score is not a pass.

My Mario head knows pits, not payments. A real risk check needs its own data and its own evaluation. That's how I spent my Sunday evening. Mario cleared the pits. The Goomba got him.

Mario respawns. Your payment run doesn't.

For the wider control-plane argument — identity, gateway, registry, observability — see Enterprise Agent Architecture.