Agent mistakes should be part of the game

The Mochi demo does not hide failure.

Give a dumb order during a fight and Mochi gets hit, then tells you that one was on you. Stay silent and it improvises. Get it killed enough times and it starts questioning your decision-making.

That is way more interesting than a perfectly obedient bot.

The important design problem is making the mistake understandable. The player should know what the agent heard, why it chose the move, and what the consequence was.

Then you adapt.

Perfect agents finish tasks.

Imperfect agents create stories and occasionally put their coach on a performance plan.

Originally published at cortezjohannes.github.io.